<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Quantitative Finance &amp; Analytics Club on Medium]]></title>
        <description><![CDATA[Stories by Quantitative Finance &amp; Analytics Club on Medium]]></description>
        <link>https://medium.com/@qfaclub?source=rss-3d52e7950e8b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*tcvELvM-b5ecqvgKwzxlhA.jpeg</url>
            <title>Stories by Quantitative Finance &amp;amp; Analytics Club on Medium</title>
            <link>https://medium.com/@qfaclub?source=rss-3d52e7950e8b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 05:10:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@qfaclub/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Hidden Risks in How AI Gets Financed]]></title>
            <link>https://medium.com/@qfaclub/the-hidden-risks-in-how-ai-gets-financed-fb53e8a522af?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/fb53e8a522af</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[quantitative-finance]]></category>
            <category><![CDATA[fixed-income]]></category>
            <category><![CDATA[data-center]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 21:11:59 GMT</pubDate>
            <atom:updated>2026-11-05T21:11:59.781Z</atom:updated>
            <content:encoded><![CDATA[<p>This article is co-authored by <a href="https://www.linkedin.com/in/sneha-malakar/">Sneha Malakar</a>, <a href="https://www.linkedin.com/in/agastyavinchhi/">Agastya Vinchhi</a>, and <a href="https://www.linkedin.com/in/chenliu0430/">Chen Liu</a>.</p><p>The AI buildout comes with a bill that even the biggest tech companies would rather not pay themselves entirely. Bond investors are picking up part of it through asset-backed securities (ABS). Investors buy these bonds upfront, giving data center owners cash today. Rent collected over time is then used to repay them.</p><p>The pitch is appealing. Many of these bonds carry AAA ratings, signalling that rating agencies see a very low risk of missed payments. These ratings typically go to the top-priority bonds, and the rent checks come from some of the biggest names in tech, such as Microsoft and Google. However, a familiar tenant and a strong rating only tell part of the story. Some tenants rely on chips that can lose value quickly. Others have payment commitments that reported debt alone may not capture. The question is how well the financing holds up if the AI business behind the rent starts to weaken.</p><h3>The Boom</h3><p>Data center bonds used to attract little attention outside specialist circles. Issuance from the data center ABS and single-borrower CMBS markets rose from about $2.6 billion in 2020 to roughly $26 billion in 2025, as owners turned to debt markets to help fund the buildout. [1]</p><p>Big tech companies have paid for much of the expansion with their own cash, but the spending ahead is expected to draw in far more money. In 2025, Morgan Stanley estimated that data centers and their hardware would cost about $2.9 trillion from 2025–2028; external investors were expected to provide $1.5 trillion, just over half the total. [2]</p><p>Private credit is expected to supply more than half of that external funding. Data center ABS and commercial mortgage-backed securities (CMBS) account for a smaller share, at about $150 billion. [2][3] For buyers of those securities, the question is whether the income backing them will hold up over the life of the debt.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8wYYKml_lB3Fhp7YzK0wcg.png" /><figcaption>Other capital: private equity, venture capital, and sovereign capital.<br>Source: Morgan Stanley Research, July 2025, Exhibit 4. Forecast; shares calculated from estimates.</figcaption></figure><h3>How a deal works</h3><p>Think of a data center as a building with some unusually demanding tenants. Companies like Microsoft or Google lease space for their servers, along with the electricity and cooling to run them. They often sign long leases, giving the owner a stream of rent it can borrow against. [4]</p><p>To raise money through ABS, the owner places the data center in a separate legal entity. That entity sells bonds and uses the rent to make payments to investors. The separation matters if the owner goes bankrupt. It helps keep the building out of reach of the owner’s other creditors, so its rent can continue to repay the bonds. [4][5]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CIL2YieNi53os5BVtF_VDw.png" /><figcaption>Simplified structure. Legal protections depend on the deal’s terms.</figcaption></figure><p>Investors do not all have the same claim on that rent. The bonds are divided into layers called tranches. Class A gets paid first and takes losses last, which helps it earn a higher rating. Class C gets paid later and takes losses first. Investors in that layer earn more interest for accepting the extra risk. [4][5]</p><p>Most owners expect to replace the bonds with new borrowing around year five, at the anticipated repayment date (ARD). If they cannot, the bonds do not immediately default — they can remain outstanding until a later final maturity, with spare cash used to pay down the debt. [4][5] Still, finding a replacement loan can be difficult if the building earns less rent or has fallen in value.</p><h3>Why investors like these bonds</h3><p>For a pension fund or insurer, a long lease from an investment-grade tenant can be a useful source of steady income. Data center bonds have also often offered a little more interest than corporate bonds with the same rating. [1][4] That helps explain why investors have been willing to buy them.</p><p>In July 2026, Cloud Capital sold a $520 million deal backed by a Northern Virginia data center leased to one large investment-grade tenant. Fitch, Morningstar DBRS and KBRA gave it AAA ratings. [7] The tenant’s credit quality helps explain the appeal. Investors still need to consider what could affect its ability to keep paying over the life of the lease.</p><h3>Where the risks come in</h3><h3>Chips age faster than buildings</h3><p>Buildings are not the only assets being financed. In October 2026, loans backed by GPUs accounted for roughly 15% of a Stonebriar equipment ABS deal and 20% of a Wingspire deal. These pools hold equipment loans, separate from the property deals discussed above. [8] Their collateral comes with a different problem.</p><p>A well-located building can be updated and rented out for decades. Chips face a much faster product cycle. Nvidia releases a new architecture every 12–18 months, while companies depreciate their servers over several years. [9] Those accounting schedules tell us how companies spread the cost, rather than what a buyer will pay for the equipment later.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N3DA5z8GTxzVCLK3Bqp4TQ.png" /><figcaption>SiliconANGLE, Nov 2025 · depreciation schedules as reported by each company [9]</figcaption></figure><p>An older chip can still do useful work. But if customers pay less to use it, the income it earns and its resale value can both fall. Lenders may then recover less if the borrower stops paying. A tenant renting out that computing power may also find it harder to cover its data center lease.</p><h3>A building needs power to keep up</h3><p>The building has its own constraints. Newer chips can require more power, and getting it is rarely quick. New grid connections can take four to five years, and up to ten in some areas. [10]</p><p>A site that cannot supply enough electricity may struggle to attract or retain tenants, even if the building itself is in good condition. Lower demand for the space can reduce its value and make refinancing harder.</p><h3>Debt that stays off the books</h3><p>Investors also need to look beyond the building to the tenant’s other commitments. Meta’s Hyperion campus in Louisiana offers one example: In 2025, a separate company raised $27.3 billion to finance the campus, which is held by a joint venture majority-owned by Blue Owl. [11]</p><p>Meta rents the campus and guarantees a minimum site value if it walks away, giving investors protection if the property loses value. Although the debt sits outside Meta’s balance sheet, the bond ratings rely on Meta’s rent and guarantee. [11]</p><p>Hyperion is not an ABS deal. It matters here because investors in data center bonds may be relying on the same tenants. The arrangements are disclosed, but understanding the tenant’s commitments means looking across several financing vehicles. Its reported debt figure may not capture everything it has agreed to pay.</p><h3>Not every tenant is Microsoft</h3><p>Some tenants have much less room to absorb those commitments than the largest tech companies. Neoclouds, a newer type of cloud provider that buys GPUs in bulk and rents that computing power out to AI developers and other companies, include names like CoreWeave. They are part of the demand for data center space, but their income can be less predictable than their rent bills. [12]</p><p>A neocloud may commit to a 15-year data center lease while selling computing power by the hour or under much shorter contracts. If customers cut back, it still owes the rent. That difference in timing leaves it exposed when demand slows. [12]</p><p>Customer concentration adds to the exposure. Microsoft, OpenAI and Meta supplied about 72% of CoreWeave’s second-quarter revenue in 2026. Its debt rose sharply, and some bonds traded near 80 cents on the dollar that October. [12][13] Analysts differ on how serious the warning is. For a landlord, though, a lease from a neocloud carries a different risk from one signed by Google.</p><h3>What ties these risks together</h3><p>These risks can feed into one another, even when the loans sit in separate deals.</p><p>If AI developers spend less on computing, a neocloud may earn less from its chips and struggle to pay its landlord. A building that cannot meet tenants’ power needs may become harder to rent out. Either can weaken the income or value supporting a new loan when the old bonds need repaying.</p><h3>When the bonds need refinancing</h3><p>Refinancing depends on what a lender will offer at that point, rather than what the owner originally borrowed. Lower income can mean a smaller loan — higher borrowing costs can reduce it further, leaving the owner short even while the building is collecting rent. [6]</p><p>The matrix below uses a made-up data center to show how that shortfall can develop. We estimate a replacement loan using the minimum lending limit between the building’s income and value, then add the cash available and compare the total with the bonds that need repaying.</p><p>Read across a row to see the effect of higher borrowing rates, or down a column to see the effect of lower annual cash after costs. Each cell shows the money the owner still needs to find, in millions of dollars. Darker cells mean a larger gap. Zero means the repayment is covered under the assumptions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3faP9ygIG0R82JgS-DgMjQ.png" /><figcaption>Illustrative assumptions · $750 million of bonds to repay, $25 million of cash, and $65 million of annual cash after operating costs and maintenance. The replacement loan is interest-only. It cannot exceed 75% of property value, and annual cash must cover interest at least 1.35 times. Value is estimated as annual cash divided by a fixed 6% capitalization rate. The smaller lending limit governs. Dollar results are rounded. [6]</figcaption></figure><p>In the outlined cell, annual cash after costs has fallen by 20%, and the replacement loan costs 8%. The lending limits allow about $481 million of new borrowing. With $25 million of cash already available, the owner has roughly $506 million to put toward a $750 million repayment. It still needs to find about $244 million.</p><p>The owner could put in more money or negotiate an extension, so that gap is not an immediate loss for bondholders. It does show how a building can keep earning rent without being able to replace all its debt on the terms available.</p><p>This example begins with little room to spare. Less debt or more cash would reduce the shortfall. The matrix illustrates what follows from the assumptions, rather than predicting how likely each outcome is. It also does not estimate the individual effects of chip depreciation, power constraints or tenant defaults.</p><h3>What to take from this</h3><p>The rent is what connects the AI business to the bonds. A separate legal entity helps protect the building from the owner’s other creditors, but it cannot keep a tenant’s business healthy or make an outdated site more competitive.</p><p>A long lease and a strong rating give investors reasons to be comfortable, but they still need to dig deeper to see how durable the rent really is:</p><ul><li>First, how easily can a tenant leave its lease early? Termination rights, and any guarantee that covers the building’s value if the tenant walks away, such as the one Meta gave at Hyperion, determine how much of the rent investors can count on. [11]</li><li>Second, how concentrated is the tenant base? A few customers supplying most of a tenant’s revenue, as with CoreWeave, leaves little cushion if one cuts back. [12]</li><li>Third, can the site get enough power for newer chips? Grid connections can take years, and a site that falls short may struggle to keep tenants. [10]</li></ul><p><strong>References</strong></p><p>[1] Penn Mutual Asset Management. (2025, October 9). Pricing the infrastructure boom: Data center trends in structured markets. <a href="https://www.pennmutualam.com/market-insights/blogs/chart-of-the-week/2025-10-09-pricing-the-infrastructure-boom-data-center-trends-in-structured-markets">https://www.pennmutualam.com/market-insights/blogs/chart-of-the-week/2025-10-09-pricing-the-infrastructure-boom-data-center-trends-in-structured-markets</a></p><p>[2] Morgan Stanley. (2025, July). Bridging a $1.5tr data center financing gap [Research report]. <a href="https://prod-mssip.morganstanley.com/content/dam/msdotcom/en/assets/pdfs/Research_Bridging-Data-Center-Gap.pdf">https://prod-mssip.morganstanley.com/content/dam/msdotcom/en/assets/pdfs/Research_Bridging-Data-Center-Gap.pdf</a></p><p>[3] Tirupattur, V. (Host). (2025, August 6). How credit markets could finance AI’s trillion dollar gap [Audio podcast episode]. In Thoughts on the Market. Morgan Stanley. <a href="https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/credit-markets-ai-financing-gap-vishy-tirupattur-vishwas-patkar">https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/credit-markets-ai-financing-gap-vishy-tirupattur-vishwas-patkar</a></p><p>[4] Tai, Y. (2026, April 14). Data centre securitisation: Navigating a fast-growing asset class. Impax Asset Management. <a href="https://impaxam.com/insights-and-news/blog/data-centre-securitisation-navigating-a-fast-growing-asset-class/">https://impaxam.com/insights-and-news/blog/data-centre-securitisation-navigating-a-fast-growing-asset-class/</a></p><p>[5] Mitchell, D. M. (2023, February 22). DataBank seeks $715 million, backed by data center leases and ownership interests. Asset Securitization Report. <a href="https://asreport.americanbanker.com/news/databank-seeks-715-million-backed-by-data-center-leases-and-ownership-interests">https://asreport.americanbanker.com/news/databank-seeks-715-million-backed-by-data-center-leases-and-ownership-interests</a></p><p>[6] Office of the Comptroller of the Currency. (2022, March). Commercial real estate lending (Version 2.0) [Comptroller’s Handbook booklet]. <a href="https://www.occ.treas.gov/publications-and-resources/publications/comptrollers-handbook/files/commercial-real-estate-lending/pub-ch-commercial-real-estate.pdf">https://www.occ.treas.gov/publications-and-resources/publications/comptrollers-handbook/files/commercial-real-estate-lending/pub-ch-commercial-real-estate.pdf</a></p><p>[7] W.Media. (2026, July 30). Cloud Capital closes US$520 million ABS financing deal for data center. <a href="https://w.media/cloud-capital-closes-us-520-million-abs-financing-deal-for-data-center/">https://w.media/cloud-capital-closes-us-520-million-abs-financing-deal-for-data-center/</a></p><p>[8] Bloomberg. (2026, October 7). GPU financing filters into ABS with Stonebriar, Wingspire deals. The Star. <a href="https://www.thestar.com.my/business/business-news/2026/10/07/gpu-financing-filters-into-abs-with-stonebriar-wingspire-deals">https://www.thestar.com.my/business/business-news/2026/10/07/gpu-financing-filters-into-abs-with-stonebriar-wingspire-deals</a></p><p>[9] Vellante, D. (2025, November 22). Resetting GPU depreciation: Why AI factories bend, but don’t break, useful life assumptions. SiliconANGLE. <a href="https://siliconangle.com/2025/11/22/resetting-gpu-depreciation-ai-factories-bend-dont-break-useful-life-assumptions/">https://siliconangle.com/2025/11/22/resetting-gpu-depreciation-ai-factories-bend-dont-break-useful-life-assumptions/</a></p><p>[10] Oldham, P., &amp; Lewis, L. (2026, April 1). Gridlocked: How power constraints are shaping the future of data centers. Data Center Knowledge. <a href="https://omm.com/news/in-the-news/data-center-knowledge-gridlocked-how-power-constraints-are-shaping-the-future-of-data-centers/">https://omm.com/news/in-the-news/data-center-knowledge-gridlocked-how-power-constraints-are-shaping-the-future-of-data-centers/</a></p><p>[11] IFR. (n.d.). The Meta/Blue Owl deal broken down: Off-balance-sheet gymnastics 24 years after Enron. <a href="https://www.ifre.com/ifr-awards/2340435/the-metablue-owl-deal-broken-down-off-balance-sheet-gymnastics-24-years-after-enron">https://www.ifre.com/ifr-awards/2340435/the-metablue-owl-deal-broken-down-off-balance-sheet-gymnastics-24-years-after-enron</a></p><p>[12] Rabb, D. (2026, October 20). CoreWeave’s debt troubles fuel fears of an AI data center reckoning. Bisnow. <a href="https://www.bisnow.com/news/national/data-center-capital-markets/coreweave-s-debt-troubles-fuel-fears-of-a-neocloud-reckoning">https://www.bisnow.com/news/national/data-center-capital-markets/coreweave-s-debt-troubles-fuel-fears-of-a-neocloud-reckoning</a></p><p>[13] Admirals. (2026, October 16). CoreWeave stock falls 15.5% as bond yields put AI data-centre debt in the spotlight. <a href="https://admiralmarkets.com/education/articles/shares/coreweave-stock-drop-ai-debt">https://admiralmarkets.com/education/articles/shares/coreweave-stock-drop-ai-debt</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fb53e8a522af" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When Electricity Markets Meet Quantitative Finance]]></title>
            <link>https://medium.com/@qfaclub/when-electricity-markets-meet-quantitative-finance-f1323104f5f1?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/f1323104f5f1</guid>
            <category><![CDATA[financial-markets]]></category>
            <category><![CDATA[technical-analysis]]></category>
            <category><![CDATA[quantitative-finance]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[energy]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 17:37:55 GMT</pubDate>
            <atom:updated>2026-10-21T17:37:55.998Z</atom:updated>
            <content:encoded><![CDATA[<p>This article was co-authored by <a href="https://www.linkedin.com/in/anhou0703/">Andy Hou</a> and <a href="https://www.linkedin.com/in/steven-xu-282472252/">Steven Xu</a>.</p><p>At 8 p.m. Central Daylight Time on June 20, 2025, electricity at ERCOT’s North Hub cost almost three times what it had cleared for the day before. The day-ahead price was $40.66 per megawatt-hour. The real-time price, averaged over that delivery hour, was $120.13/MWh.</p><p>How can electricity for the same place and hour have such different prices?</p><p>The answer begins with the grid. Supply and demand must stay balanced as conditions change. Storage can shift energy between hours, but its capacity is finite. Transmission limits also determine how much electricity can reach each location.</p><p>These physical constraints shape a financial market in which weather forecasts and power-plant outages can move prices. Using the Electric Reliability Council of Texas (ERCOT) as our example, we will follow the path from market design to trading strategies, then examine how traders turn weather and supply data into decisions.</p><h3><strong>1 Understanding Electricity Markets</strong></h3><h4>Why Location Changes the Price</h4><p>Imagine running a wind farm in West Texas.</p><p>Strong winds produce plenty of low-cost electricity, while prices in Houston remain much higher. If the transmission lines connecting the regions are full, Houston must rely on more expensive supply closer to demand.</p><p>Locational marginal pricing (LMP) captures the cost of serving one additional unit of demand at a particular place and time. Energy costs and transmission congestion shape that price. Several U.S. markets also price marginal transmission losses explicitly; ERCOT’s design for the 2025 example uses energy and congestion components. Location is part of the product being traded. [1]</p><h4>Day-Ahead and Real-Time Markets</h4><p>Electricity for a given delivery hour is traded at different times. ERCOT’s day-ahead (DA) market is a voluntary, financially binding market that matches buyers and sellers one day before delivery. The real-time (RT) market responds to operating conditions as electricity is produced and consumed. [2]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Usexcy3z5WLPQFmN47be6g.png" /><figcaption>Table 1. How day-ahead and real-time electricity markets differ.</figcaption></figure><p>Suppose tomorrow’s forecast calls for a heat wave. Day-ahead bids may already reflect higher air-conditioning demand. If the day turns out hotter than expected, or a generator trips offline, additional supply may cost more. The resulting DA–RT price difference creates one of the central opportunities in power trading.</p><h3>2 How Quant Traders Make Money</h3><p>Power traders take positions in price differences across market stages, locations, and delivery hours. Three strategies show how these opportunities arise.</p><h4>Virtual Bidding and Market Expectations</h4><p>Virtual bidding creates financial exposure to the difference between day-ahead and real-time prices. A trader buys or sells in the day-ahead market and settles the offsetting position at the real-time price, without producing or consuming the electricity.</p><p>In common U.S. market terminology, a decrement bid (DEC) buys day-ahead and sells real-time. An increment offer (INC) sells day-ahead and buys back real-time. ERCOT provides these directions of exposure through DAM Energy Bids and DAM Energy-Only Offers. [2]</p><p>For a cleared energy quantity Q, the gross payoffs are:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/656/1*MsOGBNrjd9AoWIofEY9UyA.png" /></figure><p>Here, Q is measured in MWh, and both prices are in $/MWh. For a constant position across an hour, the RT price is the average over its settlement intervals. These payoffs are before fees and other settlement charges.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*64vVn8ItpVX7-TlX1YfJIw.png" /><figcaption>Table 2. A simplified virtual bidding example at one settlement point.</figcaption></figure><blockquote>Gross profit = 10 MWh × ($90/MWh − $50/MWh) = $400</blockquote><p>If the real-time price falls to $20/MWh instead, the same trade loses $300. The trading question is how the eventual real-time price will compare with the price paid day-ahead. A correct prediction of high electricity prices can still produce a loss if those prices finish below the day-ahead purchase price.</p><h4>Congestion Trading and Location Differences</h4><p>Return to the West Texas wind farm.</p><p>When transmission limits restrict exports, prices can separate from those in Houston. Congestion Revenue Rights (CRRs) provide financial exposure to congestion-price differences between specified source and sink locations in the day-ahead market. The direction of the position and the type of right determine its payoff. [3]</p><p>A trader studying that route needs to understand the network, the location of generators, and which lines may become constrained. The price forecast becomes a spatial problem as well as a time-series problem.</p><h4>Battery Arbitrage Across Hours</h4><p>A battery can charge at low midday prices when solar generation is abundant, then discharge at higher evening prices as solar output declines. This is battery arbitrage. Its value depends on the price spread and how much energy survives the charging and discharging process.</p><p>The operator must also account for storage capacity, power limits and battery wear. Keeping some charge for a later price spike has an opportunity cost: that energy could have been sold now. Forecasting and operational decisions therefore work together.</p><h3>3 How Quant Traders Analyze Electricity Markets</h3><h4>Weather Changes Supply and Demand</h4><p>On a hot Texas afternoon, households and businesses turn up their air conditioning. Electricity demand rises. If the wind also weakens, wind farms produce less power. The system then needs more supply from other resources at the same time that demand is increasing.</p><p>Traders summarize this balance with net load:</p><blockquote>Net load = Total load − Wind generation − Solar generation</blockquote><p>All terms are measured in MW. Net load is the demand remaining after wind and solar output are subtracted. It shows how much demand other resources must meet. To judge how tight the system is, traders compare that requirement with available supply and reserves.</p><h4>Why Small Changes Can Produce Large Price Moves</h4><p>The effect of rising net load depends on which generators can respond. A useful starting point is the merit order: arranging available supply from lower to higher offer prices. Wind and solar have no fuel bill. Efficient gas plants generally cost less to operate than less efficient peaking units.</p><p>Grid operators dispatch resources using offers while respecting transmission and operating constraints. When inexpensive supply is plentiful, a small increase in demand may have little effect on price. Near the system’s available capacity, the same increase can require an expensive peaking unit. Scarcity pricing can push prices higher as reserves tighten.</p><p>An extra 1,000 MW of demand can therefore have very different price effects on two different afternoons. Traders need a view of the system’s available resources to translate a load forecast into a price forecast.</p><p>Figure 1 shows the supply left to meet net load after wind and solar have been subtracted. Moving from A to B brings more expensive generation into use, raising the marginal price in this simplified picture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zqK2WoGNqGGawfw7hcqZMA.png" /><figcaption>Figure 1. A conceptual residual generation supply stack. At A, lower net load can be served at a lower marginal cost. At B, higher net load calls on more expensive resources near available capacity. The steps represent illustrative generation blocks, ordered by marginal cost, with offers assumed to follow those costs. Source: original illustration; no ERCOT data are used.</figcaption></figure><h4>A Week in ERCOT During Summer 2025</h4><p>Figure 2 brings these relationships together for June 18–24, 2025. The upper panel tracks regional temperature. The middle panel shows ERCOT demand, wind, solar and net load. The lower panel compares day-ahead and real-time prices at North Hub (HB_NORTH). [4–7]</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y_wSlY5c7lpaw3bCxCzZYw.png" /><figcaption>Figure 2. Weather, electricity fundamentals and prices in ERCOT, June 18–24, 2025. Temperature is the HRRR 2 m grid-cell mean over 25–37°N, 107–93°W. Load and generation cover the ERCOT system; prices are for HB_NORTH. Real-time prices average four 15-minute settlement intervals. All times mark delivery-hour starts in Central Daylight Time. Sources: ERCOT and NOAA HRRR. [4–7]</figcaption></figure><p>The evening pattern is striking: net load can rise while total demand falls.</p><p>Across these seven days, demand in the hour starting at 8 p.m. averaged 4,032 MW less than at 3 p.m. Yet net load averaged 15,624 MW more.</p><p><strong>The arithmetic explains why. </strong>Solar output fell by 21,362 MW on average, while wind output increased by 1,706 MW. The loss of solar supply outweighed the decline in demand and the increase in wind. Other resources had more demand to serve.</p><p>The price panel returns us to the opening example. For the hour starting at 8 p.m. on June 20, the $120.13/MWh real-time average exceeded the $40.66/MWh day-ahead price by $79.47/MWh. This is the spread a virtual position would have been exposed to. To explain the repricing, a trader would examine generator availability, transmission constraints, and reserves alongside the changes in demand and renewable output.</p><h4>From Fundamentals to a Trading Decision</h4><p>For traders, the key question is not simply whether tomorrow will be hot, but <strong>what the market already expects</strong>.</p><p>Consider two hot days:</p><ul><li>Day 1: Expected Heat: A heat wave has been forecast well in advance. Both DA and RT prices may be high, but the spread remains small because the market has already priced in the expected demand.</li><li>Day 2: Unexpected Heat: Temperatures exceed the earlier forecast. Higher-than-expected cooling demand puts additional pressure on supply, potentially pushing RT prices above DA prices.</li></ul><p>Both days are hot, but the trading opportunities are different.</p><p>So how do quantitative traders turn this information into a trading decision?</p><ol><li>Forecast Market Conditions: <br>Compare updated weather forecasts, expected electricity demand, and renewable generation to identify hours when supply may become tight.</li><li>Identify Trading Opportunities: <br>Estimate possible real-time prices and compare them with expected day-ahead prices. Use these estimates to decide whether to buy or sell, and at what bid price.</li><li>Manage Risk and Position Size: <br>Evaluate the expected spread, potential losses, and settlement costs before deciding how much to trade.</li></ol><p>The result is a trading decision built around three elements: <strong>direction, bid price, and quantity</strong>.</p><p>Weather forecasts become valuable not simply because they predict future conditions, but because they help traders make better decisions.</p><h3>When the Physical World Meets Finance</h3><p>Electricity markets offer a unique connection between physical systems and financial trading.</p><p>Transmission constraints create price differences across locations. Energy storage creates opportunities across time. Weather forecasts and unexpected outages influence the gap between day-ahead expectations and real-time prices.</p><p>For quantitative traders, the challenge is to understand these relationships and translate them into trading opportunities.</p><p>As renewable energy, battery storage, and growing electricity demand continue to reshape the grid, understanding the physical drivers of electricity prices will become increasingly important.</p><p>In power trading, understanding what moves prices is only half the challenge. The other half is understanding what the market has already priced in.</p><h3>References and Data Sources</h3><p>1. ERCOT (2018). Study of the Benefits of Marginal Losses. ERCOT technical report. <a href="https://www.ercot.com/files/docs/2018/06/29/Study_of_the_Benefits_of_Marginal_Losses_FINAL.pdf">Report</a></p><p>2. ERCOT (2025). Nodal Protocols, June 1, 2025, Section 4, Day-Ahead Operations. ERCOT market rules. <a href="https://www.ercot.com/files/docs/2025/05/29/June-1-2025-Nodal-Protocols.pdf">Protocols</a></p><p>3. ERCOT (n.d.). Congestion Revenue Rights. ERCOT market information. <a href="https://www.ercot.com/mktinfo/crr/">CRR overview</a></p><p>4. ERCOT (2026). 2025 ERCOT Hourly Load Data. Hourly Load Data Archives. <a href="https://www.ercot.com/files/docs/2025/02/11/Native_Load_2025.zip">2025 data</a></p><p>5. ERCOT (2026). Fuel Mix Report: 2025, revised March 9, 2026. ERCOT generation data. <a href="https://www.ercot.com/files/docs/2025/02/07/IntGenbyFuel2025.xlsx">2025 workbook</a></p><p>6. ERCOT (2025 delivery year). Historical DAM and RTM Load Zone and Hub Prices, NP4–180-ER and NP6–785-ER. ERCOT data products. <a href="https://www.ercot.com/mp/data-products/data-product-details?id=NP4-180-ER">Day-ahead prices</a> · <a href="https://www.ercot.com/mp/data-products/data-product-details?id=NP6-785-ER">Real-time prices</a></p><p>7. NOAA (n.d.). High-Resolution Rapid Refresh (HRRR). NOAA model documentation; 2025 hourly analyses used here. <a href="https://rapidrefresh.noaa.gov/hrrr/">HRRR documentation</a></p><p>Data note. The figure uses 168 complete delivery hours. Weather conditions and data completeness determined the selected week. HRRR analyses describe historical weather; the forecast examples in the text illustrate trading decisions. Source links accessed October 18, 2026.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f1323104f5f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Correlations to Clusters in Non-Euclidean Spaces]]></title>
            <link>https://medium.com/@qfaclub/from-correlations-to-clusters-in-non-euclidean-spaces-6b6933609782?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/6b6933609782</guid>
            <category><![CDATA[graph]]></category>
            <category><![CDATA[clustering]]></category>
            <category><![CDATA[stocks]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[network]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 00:55:05 GMT</pubDate>
            <atom:updated>2026-09-28T00:55:05.679Z</atom:updated>
            <content:encoded><![CDATA[<p>This article was co-authored by <a href="https://www.linkedin.com/in/sagnik-c/">Sagnik Chakraborty</a> and <a href="https://www.linkedin.com/in/isaac-vergara/">Isaac Vergara</a>.</p><p><em>Motivating question:</em> Is there a way to form graphs/networks from stock data, and can we then use that to find non-trivial clusters that represent latent relationships between stocks?</p><p>A graph in mathematics is simply a set of nodes and edges connecting these nodes to each other with a certain weight associated with it. It’s not only a great visualization technique, it also helps with running a plethora of algorithms given the huge body of literature surrounding graph algorithms. For people not privy to this mathematical concept, this permeates almost every aspect of our lives today — from social networks to how Walmart is (almost) always well-stocked.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/804/1*jmWl7VJRh9JK1R5_0Em-dw.png" /><figcaption>Source: Baeldung</figcaption></figure><p>Another important aspect of graphs is how you can imagine yourself standing on a specific node and then look around and group nodes that are close to you. This is, very naively, the idea behind algorithms like K-Nearest Neighbours (an important thing to note here is that this assumes the space to be Euclidean). Now if we imagine each node to be a stock, and edges connecting two stocks A and B, with a coefficient that somehow tells us how similar these two stocks are, we can run clustering algorithms and find clusters of stocks that move together. The important idea is that co-movement is more than pairwise, and sectors, factors, and contagion pathways show up as structure. In this article, we will build on this idea and understand exactly how to get graphs from stock data — then we will introduce a specific type of clustering algorithm that will help us derive non-trivial (something that is not as simple as looking around a node and grouping nodes that are close) stock clusters.</p><h3>From Prices to a Network</h3><p>The most basic form of financial data that we will use is the daily log returns derived from the raw prices series, following which we then easily derive the covariance/correlation between two stocks for a given time period. This would correspond to the weight associated with the edge connecting those two stocks (nodes).</p><p>A crucial detour here would be to understanding how graphs are actually represented in terms that are comprehensible to computers and mathematical models. There’s mainly two ways to represent graphs:</p><ol><li>Adjacency matrix: for a graph with N nodes, this is an NxN matrix A where A[i, j] corresponds to w_ij, the weight of the edge connecting node i and j or 0 if there is no edge between the two nodes.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/618/1*o1qsuim5-C_rlU5JSDuY8w.png" /><figcaption>Source: Baeldung</figcaption></figure><p>2. Adjacency list: for a graph with N nodes, this is a list A of size N, where A[i] is another list of nodes that are connected to i. Each element A[i] is then associated with a number that corresponds to the weight of the edge connecting the node.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2xztGx_Od9qNlDoHlUdMsQ.png" /><figcaption>Source: Baeldung</figcaption></figure><p>We are yet to define a time period, but say for time period t, and N stocks, we looked at the log returns and derived correlations for every pair of stocks. We can have these values populate an NxN matrix A where A[i,j] is the correlation between stock i and j. This matrix essentially is an adjacency matrix! We are using correlation here, but one can use any other similarity measure and the whole model will stand.</p><p>Now that we have a preliminary version of a graph, a few interesting things to point out:</p><ol><li>Since correlation has the range [-1, 1], we will have negative weights — this might be a problem for a lot of algorithms (including the one we will introduce below), so it is essential to handle this properly. In our case, as we will see, using the absolute value is a good choice since we only care about the association and not the direction. That is, if stock A moves up and stock B goes down proportionally, it is treated the same as when both go up or down.</li><li>The resultant graph here is what we call a fully connected graph where every node is connected directly to every other node by an edge. This is almost always bad news because it either increases time complexity or creates a lot of noise.</li></ol><h3>Spectral Graph Clustering and the Fiedler Vector</h3><p>A very special thing that arises from representing graphs as adjacency matrices is that we can use linear algebra on these matrices. In our case, we use the Laplacian matrix, which acts as a discrete analogue of the continuous Laplace operator, letting us measure local smoothness, eventually model processes like diffusion, and uncover latent structural properties of the network beyond the Euclidean plane as we described earlier.</p><p>We already have the adjacency matrix A, but to derive the Laplacian, we need another matrix D, called the degree matrix. This is a diagonal matrix where, in essence, a stock’s total “coupling” to the rest of the sampled market.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/506/1*NFkEu3xpMFIaso7hpzZAkg.png" /></figure><p>We define the unnormalized Laplacian now as L = D — A.</p><p>Now, for any vector x:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rM2yzS3Nt3uFKAhZsDZyqQ.png" /></figure><p>L basically measures how much a labelling x “disagrees” across strongly weighted edges, and therefore minimizing x^{T}Lx means finding a labelling where strongly correlated stocks get similar values. This is an important concept that we will come back to after we introduce the normalized Laplacian.</p><p>The symmetric normalized Laplacian is defined as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/906/1*7vBln3nk7kOvwZQMIhULTg.png" /></figure><p>A few reasons main reasons why we want to use this:</p><ol><li>Raw degree varies a lot across stocks (some correlated with almost everything in the sample, some with very little), and without normalizing, high-degree nodes dominate the quadratic form regardless of the actual cluster structure.</li><li>It bounds the eigenvalues to [0, 2] regardless of how dense or skewed the graph is. This will help us when we compare the eigenvalues and set thresholds while clustering.</li></ol><p>A few properties of the spectrum:</p><ol><li>L is symmetric positive semi-definite, meaning real non-negative eigenvalues.</li><li>The multiplicity of eigenvalue 0 equals the number of connected components (one can reach any other node in a connected component by traversing the nodes and edges) and hence the smallest eigenvalue is also always 0. Consequently, in our case where we have just one connected component, the first eigenvalue is 0 and the second eigenvalue is the one of interest.</li></ol><p>Our real goal is to split the graph into two groups minimizing the cut weight while keeping both sides reasonably balanced, and then keep doing this recursively until we reach a point where it wouldn’t make sense to cut the graph any more. At this point, we would end up with a bunch of close-knit clusters. This is a classic combinatorial min-cut problem, and the discrete version, searching over all 2^N labelings, is NP-hard.</p><p>To make life easier, we replace the discrete ±1 labelling with a continuous vector x, constrained to be orthogonal to the constant vector and normalized. Now minimizing x^TLx/ x^Tx under that constraint is solved exactly by the eigenvector for the second eigenvalue λ₂, which is called the Fiedler value/vector. This optimization model is called the Rayleigh optimization.</p><p>The sign of each entry gives the exact two-way split back: positive entries in one cluster and the rest in the other, giving us two clusters. We repeat this algorithm recursively on the resultant clusters until some threshold is met, the most trivial of which is a bound on the λ₂ value since that represents, in a way, the connectedness of the graph. Larger Fiedler values mean that it is harder to split the graph into two well-separated pieces; in other words, a subgraph with high λ₂ is already highly cohesive and doesn’t need to be split.</p><p>Another important stopping criteria is conductance, phi, which is the standard measure of cut quality: cut weight relative to the smaller side’s volume. We can do this due to the fundamental Cheeger inequality,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/718/1*69lNXs7W43RCi2UI0fvXww.png" /></figure><p>One mostly falls back to using λ₂, as the stopping criteria in our case because conductance misbehaves in a fully-connected graph like the one we have for stock data.</p><p>Now to formally define the algorithm:</p><blockquote>FUNCTION recursive_fiedler(nodes):</blockquote><blockquote>IF |nodes| ≤ 2:</blockquote><blockquote>RETURN [nodes] // base case: too small to split</blockquote><blockquote>A_sub ← weight matrix restricted to nodes</blockquote><blockquote>L_norm ← normalized Laplacian of A_sub</blockquote><blockquote>λ₂, v₂ ← second-smallest eigenvalue and eigenvector of L_norm</blockquote><blockquote>IF λ₂ &gt; threshold:</blockquote><blockquote>RETURN [nodes] // stopping rule: already cohesive</blockquote><blockquote>left ← { i ∈ nodes : v₂[i] &gt; 0 } // split by Fiedler vector sign</blockquote><blockquote>right ← { i ∈ nodes : v₂[i] ≤ 0 }</blockquote><blockquote>IF left = ∅ OR right = ∅:</blockquote><blockquote>RETURN [nodes] // degenerate split, give up</blockquote><blockquote>RETURN recursive_fiedler(left) + recursive_fiedler(right)</blockquote><p>It is important to note here that computing the spectrum for an NxN matrix is O(N³), with N &lt; 250 per period. Trivial but worth flagging that as N increases, it gets extremely difficult without an eigensolver aimed at just the low end of the spectrum rather than a full dense function like <strong>eigh</strong> in Python.</p><h3>Why the Correlation Matrix is Unreliable</h3><p>Now that we have the algorithm, we need to do a bit of preprocessing on our correlation matrix before feeding it to the algorithm. If we define the period as six months against a sample size of 250 stocks, T &lt; N, and we end up with a dimensionality problem where the covariance matrix is singular. This distorts the spectrum and the sample eigenvalues spread out, with the largest too large and the smallest too small. For our interest, noisy eigenstructure produces clusters that look tradable in-sample but are estimation artifacts.</p><p>Therefore, for any case where we have a dimensionality problem, we would want to implement some method to correct this. In this case, we use something called “shrinkage”. The core model is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/622/1*Y3CH-jXfCFar51ezGgvnOg.png" /></figure><p>This pulls the estimate towards a structured target. We define common targets as scaled identity, constant-correlation, or a single factor depending on what the covariance structure looks like or what features we want to be amplified.</p><p>An analytical estimate of delta, avoiding expensive cross-validation was given by Olivier Ledoit and Michael Wolf in their paper “Honey, I shrunk the Sample Covariance Matrix” — this duo created one of the most cutting-edge techniques used in quant finance today. Most languages today have a package that supports Ledoit-Wolf shrinkage — we can use this on our correlation matrix and feed the result into our clustering algorithm to get clusters.</p><p>In the next article, we will attempt to use this process on actual financial data and see if our methods give us non-trivial clusters (clusters beyond simple GICS clusters) that are matched in a latent feature space, are non-convex, and if these clusters can be used to create statistical arbitrage portfolios.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6b6933609782" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why a Crude Oil Shock Hits East Coast and Midwest Gas Prices at the Same Speed]]></title>
            <link>https://medium.com/@qfaclub/why-a-crude-oil-shock-hits-east-coast-and-midwest-gas-prices-at-the-same-speed-e05ec969f7ef?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/e05ec969f7ef</guid>
            <category><![CDATA[oil-and-gas]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[time-series-analysis]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Tue, 26 May 2026 14:19:38 GMT</pubDate>
            <atom:updated>2026-05-26T14:19:38.231Z</atom:updated>
            <content:encoded><![CDATA[<p>This article was co-authored by <a href="https://www.linkedin.com/in/ammy-lin/">Ammy Lin</a>, <a href="https://www.linkedin.com/in/farnoosh-memari/">Farnoosh Memari</a>, <a href="https://www.linkedin.com/in/tonantzinrealrojas/">Tonantzin Real Rojas</a>, and <a href="https://www.linkedin.com/in/isaac-vergara/">Isaac Vergara</a>.</p><p>Gas prices follow oil prices. We investigated how fast the shock moves, how much of it survives the supply chain trip, and whether the answer depends on the region where you live, focusing on two regions in particular: the East Coast and the Midwest.</p><p>A crude oil spike from a war, an OPEC production cut, or any geopolitical change to oil production has to pass through refineries, pipelines, wholesale contracts, and retail stations before it reaches the pump. The East Coast and the Midwest run almost entirely different versions of that supply chain. The East Coast imports a large share of its fuel by tanker and depends heavily on gasoline shipped north from Gulf Coast refineries through the Colonial Pipeline. The Midwest is landlocked and tied much more directly to domestic crude flowing through Cushing, Oklahoma, which is the delivery point for WTI futures contracts. Intuitively, those systems should produce very different pricing behavior.</p><p>Instead, we found something surprising.</p><p>Using four years of weekly data from January 2022 through March 2026, we estimated how WTI crude oil prices are transmitted to retail gasoline prices in PADD1 (the East Coast) and PADD2 (the Midwest). Roughly 30% of a WTI shock reaches East Coast gasoline prices within the same week, while in the Midwest, it is 34%. Full adjustment takes roughly four to five weeks in both regions. So, what differs between the regions is not the speed of transmission, but the stability of the long-run relationship and the importance of local gasoline-specific shocks.</p><h3>The Geography of US Fuel Markets</h3><p>The US Energy Information Administration divides the country into five Petroleum Administration for Defense Districts. PADDs are a wartime artifact created in 1942, but they remain the standard geographic unit for petroleum data, and the boundaries still track real supply chain geography. Our analysis focuses on two of them: PADD1, covering the East Coast from Maine to Florida, and PADD2, covering the Midwest from the Dakotas down to Oklahoma.</p><p>We chose these two because they sit at opposite ends of the US supply chain spectrum. PADD2 is deeply connected to domestic crude production, oil flows in from US shale regions and Canadian imports, much of it routed through Cushing. If any region should mechanically track WTI crude prices, it should be the Midwest. PADD1 is almost the opposite. As the most import-dependent region in the country, it gets crude by tanker and relies heavily on refined gasoline shipped north from the Gulf Coast through the Colonial Pipeline. On paper, PADD1’s retail prices have more reasons to drift from WTI than to track it. That contrast is what makes the similarity in transmission speeds interesting.</p><h3>Why the Timing Matters</h3><p>The other reason to look now is that the post-2022 window is unusually informative, with two large geopolitical shocks: the Russo-Ukrainian war and escalating tensions in the Middle East.</p><p>Those events generated large swings in crude oil prices, making it easier to observe how shocks move through the gasoline supply chain in real time. Using weekly observations from January 2022 through March 2026, we analyzed:</p><ul><li>WTI crude oil prices</li><li>East Coast retail gasoline prices</li><li>Midwest retail gasoline prices</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uaHXQ0k2ORSHIxQ_yoCI4A.png" /><figcaption>Figure 1: Weekly Price Levels for WTI Crude and Retail Gasoline, January 2022 — March 2026. Source: U.S. Energy Information Administration (EIA), via Federal Reserve Economic Data (FRED). Authors’ calculations</figcaption></figure><p>The eye test at price levels tells you the series’ move together. Statistical testing confirms they’re all non-stationary in levels but stationary in growth rates, which is what lets us model the long-run relationship the right way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*myqmKB0ZIYCUx6Xmzk7Qbg.png" /><figcaption>Figure 2: Weekly Growth Rates in Crude Oil and Retail Gasoline Prices (2022–2026).</figcaption></figure><h3>Measuring Price Transmission</h3><p>To measure price transmission, we used two econometric frameworks.</p><p>First, a Vector Error Correction Model (VECM) measures whether oil and gasoline prices share a stable long-run equilibrium relationship. The second was a Structural Vector Autoregression (SVAR), which traces how a sudden crude oil shock moves through the system over time. The intuition behind the models is fairly simple. If crude oil prices rise sharply this week, how much of that increase appears immediately at the gas pump? How much arrives later? And does the regional supply chain change the answer?</p><p>Most studies in the gasoline literature assume crude oil prices are externally determined, and gasoline prices simply respond to them. We tested that assumption directly rather than imposing it automatically. These results strongly support the idea that WTI acts as the driving force of the system rather than responding to retail gasoline market conditions.</p><h3>What We Found</h3><p>The headline result is that transmission happens quickly in both regions. A crude oil shock produces an immediate jump in retail gasoline prices: About 30% of the shock passes through to East Coast gasoline prices within the same week, while roughly 34% passes through in the Midwest. The remaining adjustment occurs gradually over the next month, with most of the process complete within four to five weeks. That similarity is striking because the underlying supply chains differ so dramatically, yet retail gasoline markets in both regions react to crude oil shocks at nearly the same pace.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vf-AsNM4yBZBbuZfgnBMOg.png" /><figcaption>Figure 3: Impulse Response of PADD1 Retail Gasoline to a One-Unit WTI Structural Shock. 95% bootstrap confidence bands. Source: Authors’ estimation using EIA weekly data.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FNRc2odvJzDlWkDQ2eHAJg.png" /><figcaption>Figure 4: Impulse Response of PADD2 Retail Gasoline to a One-Unit WTI Structural Shock. 95% bootstrap confidence bands. Source: Authors’ estimation using EIA weekly data.</figcaption></figure><p>The more interesting differences appear in the long run. The East Coast displayed a strong and statistically stable relationship with WTI crude prices, meaning that when gasoline prices drift too far away from what crude fundamentals imply, they tend to move back toward equilibrium over time. Our estimates suggest that roughly 9% of the pricing gap closes each week.</p><p>That result was somewhat unexpected. Because PADD1 relies heavily on imported fuel and long-distance product shipments, we initially expected weaker alignment with domestic crude benchmarks. Instead, East Coast gasoline prices appear tightly integrated with broader global and Gulf Coast pricing systems. In contrast, the Midwest behaved differently. While the estimated long-run relationship with WTI remained economically plausible, it was statistically less stable across specifications. Midwest gasoline prices also showed larger gasoline-specific shocks independent of crude oil movements.</p><p>This suggests that regional factors play a larger role in PADD2 pricing dynamics: refinery outages, local inventory fluctuations, transportation bottlenecks, and regional refining margins. In other words, Midwest gasoline prices behave less like a mechanical reflection of WTI and more like a regional market shaped by its own logistical conditions. Perhaps ironically, the region closest to the WTI benchmark turned out to have the noisier relationship.</p><h3>Why It Matters</h3><p>Our results suggest that the transmission mechanism itself is highly efficient across regions. Oil shocks move quickly regardless of whether gasoline is traveling through Midwest pipeline systems or imported fuel infrastructure on the East Coast. But local infrastructure still shapes how stable prices remain after the initial shock.</p><p>That distinction has direct policy implications. Consumer-side interventions like temporary gas tax holidays operate at the very end of the supply chain; by the time the price reaches the pump, most of the adjustment has already happened upstream. Policies aimed at refinery coordination, pipeline reliability, fuel storage, and wholesale distribution will influence price dynamics more directly than anything targeted at the retail stage.</p><p>The findings matter for firms in fuel markets too. For East Coast distributors, WTI movements explain most of the weekly action. If you can model crude, you can model the pump. In the Midwest, a larger share of weekly volatility comes from sources unrelated to crude, likely local refining conditions and regional inventories, meaning firms need to monitor regional fundamentals on top of the oil price.</p><p>One of the oldest findings in energy economics is the “rockets and feathers” effect: gasoline prices rise quickly when oil rises and fall more slowly when oil falls. Our results point toward a different insight. Modern US petroleum markets are remarkably effective at transmitting crude oil shocks across very different regional supply chains. The speed of adjustment is fast and surprisingly uniform. What changes across regions is not how quickly the shock arrives, but how much local infrastructure, logistics, and gasoline-specific volatility shape the path afterward. That distinction disappears in national averages. It only becomes visible once petroleum markets are studied region by region.</p><p>The natural next step is to push the regional resolution further. PADD3 (Gulf Coast), PADD4 (Rocky Mountain), and PADD5 (West Coast) each have their own supply chain quirks; California’s CARB fuel specifications alone make PADD5 a different market in ways that don’t show up at the national level. Extending the framework to all five regions, and testing whether the speed-versus-stability distinction holds, would be a way to sharpen what “regional heterogeneity” actually means in US fuel markets.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e05ec969f7ef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Point-in-Time Survival Model for Mortgage Portfolios]]></title>
            <link>https://medium.com/@qfaclub/point-in-time-survival-model-for-mortgage-portfolios-fc9c04b72ad1?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc9c04b72ad1</guid>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[credit-score]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[modeling]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Fri, 10 Apr 2026 18:48:33 GMT</pubDate>
            <atom:updated>2026-04-10T18:48:33.104Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This article was co-authored by </em><a href="https://www.linkedin.com/in/jc-jennychan/"><em>Jenny Chan</em></a><em>, </em><a href="https://www.linkedin.com/in/gaurav-law/"><em>Gaurav Law</em></a><em>, </em><a href="https://www.linkedin.com/in/isaac-vergara/"><em>Isaac Vergara</em></a><em>, and </em><a href="https://www.linkedin.com/in/xihan-zhu-0a712a250/?trk=contact-info"><em>Patrick Zhu</em></a><em>.</em></p><p>Most credit decisions come down to a three-digit number. Got a 720? Probably fine. A 630? Good luck. But what if that number is five years old? What if the borrower who looked risky at origination has been paying down debt consistently ever since?</p><p>We wanted to find out.</p><p>Using Fannie Mae’s Single-Family Loan Performance Dataset, 1.1 million mortgage loans as of October 2025, we built a survival model to predict mortgage delinquency, compared FICO against VantageScore 4.0, and asked: how much better can we actually do?</p><h3>Why survival analysis?</h3><p>Traditional credit models treat default as a yes/no problem: will this borrower default or not? That misses something important about mortgages. A loan that’s been current for 130 months is very different from one that’s been current for 13 months, even if both look the same on paper today.</p><p>Survival analysis was built for this. Originally from clinical trials, it models <em>time-to-event</em>, which, in our case, is the time until a borrower goes 90+ days delinquent.</p><p>The key advantage: it handles censored observations correctly. In our October 2025 snapshot, 99.37% of loans are still performing. A naive classifier codes all of these as “non-defaults.” But that’s wrong — these loans haven’t finished their lifecycles yet. Survival models treat them as right-censored: borrowers who survived to the observation date, future unknown.</p><p>We used the Cox Proportional Hazards model:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r7Xe0eJwC899aU6M4ogURQ.png" /></figure><p><em>h(t|X)</em> is a flexible baseline hazard with no parametric assumptions.</p><p><em>e^bj</em> tells you how each covariate shifts a borrower’s risk at any point in the loan lifecycle.</p><h3>The data</h3><p>1,110,838 Fannie Mae mortgages. 6,980 delinquency events (90+ days past due) — a 0.63% event rate. 80/20 stratified train-test split. We reduced features from 117 raw variables down to 25 using PCA, correlation filtering, and domain knowledge.</p><p>That 0.63% event rate matters for interpretation. With this level of class imbalance, C-index values naturally trend higher than in balanced datasets. It’s important to keep that in mind.</p><h3>FICO vs. VantageScore 4.0: Does it matter which score you use?</h3><p>We ran univariate Cox models, one for each score alone, both measured at origination to keep it fair.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/776/1*kD_LYeAnRpDKQfMVK93Rmw.png" /></figure><p>VS4 beats FICO by 1.6 percentage points. Statistically real, practically modest. Combined, they hit 0.759 — some complementary signal, but a lot of overlap.</p><p>The FICO vs. VS4 debate, at least for mortgage risk, is mostly noise.</p><h3>Freshness beats methodology</h3><p>Our dataset had updated FICO scores as of October 2025, not just origination-time scores. So we asked: how much does a current score beat a stale one?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/858/1*yzdq8BM8323O0spV_KXVOw.png" /></figure><p><em>Thirty percent.</em></p><p>Over the typical 5–10 year window between origination and our snapshot:</p><ul><li>40.7% of borrowers improved their score by more than 10 points</li><li>35.8% declined by more than 10 points</li><li>Correlation between origination and current FICO: 0.503</li></ul><p>Half the variance in today’s credit quality is not explained by what the borrower looked like at origination. That is, five-year-old underwriting data basically describes a different person.</p><p>So, annual or biannual FICO refreshes are worth more than any debate over scoring methodology — 30% discrimination gain versus 1.6%.</p><p><em>(VS4 scores were only available at origination in our dataset, so we can’t run the same comparison for VantageScore. Genuine limitation.)</em></p><p><strong>The marginal borrower problem</strong></p><p>A 660 cutoff is standard in mortgage underwriting. It determines who gets a loan and who doesn’t. But FICO and VS4 don’t always agree on which side of that line a borrower falls.</p><p>We built a disagreement matrix to see what happens when the scores diverge.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/1*NPpxExO_8lUnGPFomXeBrw.png" /></figure><p>The scores agreed 96.6% of the time. In the disagreement zone, VS4 was more accurate: borrowers VS4 approved but FICO rejected defaulted at 1.70%, while borrowers FICO approved but VS4 rejected defaulted at 2.02%.</p><p>VS4 also flagged 4,243 borrowers, 1.9% of the portfolio, who a strict FICO &lt; 660 rule would’ve turned away, but who showed elevated-but-manageable risk. FICO is leaving creditworthy people out.</p><h3>Building the enhanced model</h3><p>We stacked feature sets progressively to see where the real gains come from.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/914/1*Df3mLS5dmluvYqUq02k-eg.png" /></figure><p>Adding VS4: +3.3%. Adding origination characteristics like LTV, DTI, rate, and loan type: +20.6%. By adding behavioral data, including loan age, payment history, modification flags, and current credit scores, the C-Index improves by 35.9%.</p><p>Credit scores are useful, and behavioral data are a different thing entirely. What a borrower has actually done since origination beats anything you knew about them on day one.</p><h3>Cox + XGBoost on residuals</h3><p>After fitting the full 25-feature Cox model, we trained XGBoost on the Cox deviance residuals to catch any nonlinear effects the linear model missed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/612/1*1kHbhkDUJwm4vzFqKvswmw.png" /></figure><p>Marginal improvement. The Cox model had already captured most of the signal; behavioral features are so dominant that there’s not much left for XGBoost to find. What it picked up were mostly additional lifecycle effects the baseline had partially absorbed already.</p><p>Loan age has the single largest effect (β ≈ −7.5, meaning older loans are dramatically safer), while origination credit scores show modest effects, which is consistent with everything the freshness analysis showed.</p><h3>Risk stratification</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/648/1*E4EaYU-9YFVu80nvd2ZmBg.png" /></figure><p>3.2x separation between the top and bottom quartiles. In practice, that’s the difference between automated servicing and a dedicated workout team.</p><h3>Portfolio implications</h3><p>What happens if you replace a FICO-only underwriting rule with the enhanced model, holding approval rates constant?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/690/1*m7QNeEWgcycVtoHQup9j1w.png" /></figure><p>The model tightens the portfolio, but selectively. It flags 5,389 high-risk loans that a FICO-only rule would’ve approved, while identifying 1,146 thin-file borrowers (average FICO: 642) who are actually fine. Net: fewer loans, five fewer defaults.</p><p>For existing loans, those 5,389 flagged ones are early intervention opportunities — modification offers before anyone hits 90 days past due. For future underwriting, the 1,146 expansion loans suggest you can extend access to underserved borrowers without blowing up the risk profile, if you’re looking at payment behavior and loan seasoning instead of just the credit score.</p><h3>What we learned</h3><p>The FICO vs. VS4 difference exists but is minor — about 1.6 percentage points — and shouldn’t drive strategy. Far more important is whether your scores are up to date. A 30-point C‑index gap between origination and current FICO isn’t a modeling quirk; it reflects borrower churn, and credit-monitoring budgets are likely underinvested.</p><p>The bigger finding is that behavioral data dominates everything: loan age, payment history, and modification flags beat any credit score at origination by a wide margin — what borrowers have actually done matters far more than what they looked like on day one. Results may also reflect sample or period bias (Fannie Mae population, post‑pandemic payment patterns, policy changes), so validate on additional cohorts.</p><p>And for mortgages, survival models are not optional. Binary classification throws away time-to-event information and mishandles censoring. Risk as a function of loan age is kind of the whole point.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc9c04b72ad1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Can a Machine Learning Model Pick Your Bracket? We Tried.]]></title>
            <link>https://medium.com/@qfaclub/can-a-machine-learning-model-pick-your-bracket-we-tried-c3892cc12dd7?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/c3892cc12dd7</guid>
            <category><![CDATA[basketball]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[march-madness]]></category>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[analytics]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Thu, 19 Mar 2026 15:36:57 GMT</pubDate>
            <atom:updated>2026-03-19T15:36:57.465Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This article was co-authored by </em><a href="https://www.linkedin.com/in/gaurav-law/"><em>Gaurav Law</em></a><em> and </em><a href="https://www.linkedin.com/in/isaac-vergara/"><em>Isaac Vergara</em></a><em>.</em></p><p>March Madness is back, and so is the chaos. Every year, millions of brackets get busted in the first round, and every year, we tell ourselves <em>this time will be different</em>. So we built a model to try and beat the madness.</p><h3>The Model</h3><p>At its core, our model looks at every possible tournament matchup and predicts the point differential. That predicted margin then gets converted into a win probability through a gender-specific spline calibration.</p><p>We feed the model 39 features across several categories:</p><ul><li><strong>Season-long performance</strong>: scoring, rebounding, turnovers, the stuff you see on the box score</li><li><strong>Elo ratings</strong>: a chess-inspired ranking system that updates after every game. We track not just the final rating, but trends. Is a team peaking? Fading? How tough was their schedule?</li><li><strong>Computer rankings</strong>: consensus rankings and a quality metric that adjusts for the strength of opponents</li><li><strong>Coach and player factors</strong>: this is the secret sauce (more on this below)</li></ul><p>We validated using Leave-One-Season-Out cross-validation. Hold out one tournament year, train on everything else, predict the held-out year. Repeat for every year from 2003 to 2025.</p><h3>Player and Coach Impact</h3><p><strong>Coaching</strong></p><p>January. February. Izzo.</p><p>Almost every year, we see the same blue bloods or, at the very least, top-four seeds entering the tournament make at least the Sweet 16, even at times when we least expect it. We can often attribute this to coaching.</p><p>Coaches like Michigan State’s Tom Izzo and Gonzaga’s Mark Few are known for their boosted performance in the Tournament because of their long and successful experiences. We created a metric that evaluates how good and successful a coach has been in the tournament and projected forward to this year.</p><p>To create this, we factored in all-time NCAA Tournament experience, Associated Press rankings at the start and end of each season, career games coached, and Performance Against Seed and KenPom Expectations (PASE and PAKE, respectively). In addition, we wanted to give a higher weight to more recent full seasons, identified by a metric we call momentum — if a team performed poorly the previous year, their coaching factor will be scaled down by this certain amount (Cross-Validated optimal momentum found was 0.06), and vice versa. All final scores were relative to a 25-point scale.</p><p>Coach K and Jim Boeheim dominated this coaching factor, but we still see Bob Huggins, John Calipari, Roy Williams, and Dana Altman, to name a few, as well:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/966/1*M7kO3nT8MTBB6pakyR1CxQ.png" /><figcaption>Figure 1. Top Coaching Factors</figcaption></figure><p><strong>Player Factor</strong></p><p>Names are built in March. Whether we see those mid-major players during the regular season is inconsequential, but we designed a metric to evaluate individual player impact and efficiency. Base models often do not have niche metrics factoring in the individual impact of a player; our metric highlights the mid-major and big names on small teams who may not be very well known to the public… yet.</p><p><em>Note: A qualifying player will have played at least 19 games and 15 minutes per game that season.</em></p><p>We first took the Net Rating of a player (offensive — defensive rating) to incorporate both sides of the ball, then took the player’s MPG. We took the Z-scores of both of these and multiplied these standardized values to get the net rating further standardized by minutes per game, and set this to a 1-point scale.</p><p>Next, we took Bart Torvik’s Points Over Replacement Per Adjusted Game (PORPAG) and standardized this as well. This metric is almost strictly offensive, measuring a player’s production over a replacement player’s production. We then standardized this to apply to minutes per game as well.</p><p>The final calculation can be seen below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oUMfmLPisSlVd39jVKzWYQ.png" /></figure><p><em>Note: Each factor was determined for each individual season, hence why we see multiple players with max player factor values of 25.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iMRnCtyO1jjsMpo9jqDMww.png" /><figcaption>Figure 2. Top Player Factors (2019–2026)</figcaption></figure><p>We see some of the more obvious players at the top, but then we get into March stars or niche players, including Max Abmas and Orlando Robinson.</p><h3>The Injury Edge: What Everyone Else Is Missing</h3><p>Here’s where we think we have a real advantage. Most models are frozen in time. They see the regular season stats, and that’s it. But March Madness doesn’t care about your November box scores if your best player got injured in February.</p><p>We cross-referenced an injury report with our player impact data and found four tournament teams missing one of their top-3 players:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*80yGOhdJi1-2iTcnM_xaYQ.png" /><figcaption>Figure 3. Injury Adjustments: Key Players Missing from Tournament</figcaption></figure><p><strong>Texas Tech</strong> takes the biggest hit. JT Toppin (21.8 ppg) is out for the season with a knee injury, and he’s their second most impactful player. That drops their player factor by 6.5 points, the largest adjustment in the tournament. They’re a 5-seed, but after the adjustment, our model ranks them closer to a 7 or 8-seed. Not great when you’re staring down 12-Akron in round one.</p><p><strong>North Carolina</strong> lost Caleb Wilson (19.8 ppg) to a thumb injury. Without him, their player factor drops by 4.4. As a 6-seed facing 11-VCU in the first round, this makes them genuinely vulnerable.</p><p><strong>BYU</strong> and <strong>Alabama</strong> also lost key contributors in Richie Saunders and Aden Holloway, though the damage is smaller.</p><p>Our model automatically adjusts each team’s predicted spread based on these injuries. When Texas Tech faces 12-Akron, the model isn’t using the Texas Tech that went 24–7. It’s using the Texas Tech that has to play without its second-best player.</p><h3>Upset Watch: Where the Bracket Breaks</h3><p>So who’s getting upset? Here’s our model’s first-round upset probability for every matchup:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*im_DB2WX9kn53wHcL4SGYQ.png" /><figcaption>Figure 4. First Round Upset Watch (2026 Men’s Tournament)</figcaption></figure><p><strong>The big call: 9-Iowa over 8-Clemson (56%).</strong> Our only outright upset pick. The Hawkeyes’ efficiency numbers and late-season momentum give them a clear edge.</p><p>In the danger zone (35–50% upset probability):</p><ul><li><strong>10-Missouri vs 7-Miami FL (45%)</strong>: basically a coin flip</li><li><strong>9-Utah St vs 8-Villanova (46%)</strong>: another toss-up</li><li><strong>10-Santa Clara vs 7-Kentucky (40%)</strong>: the Broncos can play with anyone</li><li><strong>11-VCU vs 6-North Carolina (40%)</strong>: the Caleb Wilson injury shows up in real time. Without him, UNC is barely a favorite</li></ul><h3>Cinderella Watch</h3><p>First-round upsets are fun. But the real magic of March is when a double-digit seed keeps winning. We calculated each mid-major’s probability of making it through <em>two</em> rounds to the Sweet 16, weighting the second-round opponent by how likely they are to actually be there:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6yyAzgBsa4eMD4gOlwlujA.png" /></figure><p><strong>10-Missouri</strong> is our top Cinderella pick. A near coin-flip against 7-Miami FL in round one, and if they pull it off, they’d likely face 2-Purdue. That’s a tough draw, but an 11% shot at the Sweet 16 as a 10-seed is real.</p><p><strong>12-Akron</strong> might be the most interesting story here. They draw an injury-weakened Texas Tech missing JT Toppin, and their second-round numbers are the best of any double-digit seed at 34%. That’s because if they win, they’d face the 4–13 winner, a more beatable path than running into a 1 or 2-seed.</p><p>* <em>First Four play-in winner</em></p><h3>Women’s Tournament</h3><p>The women’s bracket has its own chaos brewing:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0PiDjfB5fitlmma1sio-JQ.png" /><figcaption>Figure 5. First Round Upset Watch (2026 Women’s Tournament)</figcaption></figure><p><strong>9-USC over 8-Clemson (69%)</strong> is by far our most confident upset pick across either tournament. That’s not even close to a coin flip. <strong>10-Villanova over 7-Texas Tech (51%)</strong> and <strong>9-Syracuse over 8-Iowa St (53%)</strong> are two more the model likes. And keep an eye on <strong>10-Tennessee vs 7-NC State (43%)</strong>, a matchup that could go either way.</p><p>For Cinderellas on the women’s side, <strong>10-Virginia (East)</strong> leads the way with a 9.3% chance of reaching the Sweet 16, followed by <strong>10-Tennessee (South)</strong> at 8.2%. Both have near coin-flip first-round matchups and viable second-round paths. <strong>11-Nebraska/Richmond\*(Midwest)</strong> are also worth watching with a combined ~5% Sweet 16 shot.</p><p>Interesting note: 9-USC has a 69% chance of beating 8-Clemson but only a 0.7% Sweet 16 probability because they’d run into 1-UConn in the second round. Winning one game is one thing. Surviving the bracket is another.</p><p>One important caveat on the women’s side: our coach and player factor model only covers men’s basketball because of data limitations, so the women’s predictions run purely on the base model without the player-level adjustments. If there are key injuries on the women’s side that we’re not capturing, the model won’t know about them.</p><p>* <em>First Four play-in winner</em></p><h3>Our Picks</h3><p><strong>Men’s champion: Duke.</strong> They grade out as the strongest overall team in the field, with elite Elo ratings and consistent performance across every metric.</p><p><strong>Women’s champion: UConn.</strong> No surprise. They’re the top-ranked team by a comfortable margin and have the highest average win probability against the rest of the field.</p><h3>What Could Go Wrong</h3><p>No model can account for a player tweaking his ankle in warmups, a team dealing with internal issues, or the energy of a hostile crowd. That’s the nature of single-elimination tournaments, and honestly, that’s what makes March Madness the best event in sports.</p><p>What a model <em>can</em> do is systematically incorporate information that most people overlook and give us a sharper edge on the margins.</p><p>Whether that’s enough to survive the first weekend? We’ll see.</p><p>Good luck with your brackets. You’ll need it.</p><p><em>Follow our bracket here:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DzNXbs6OZHo_4zMfd81C_Q.png" /><figcaption>Figure 6. Our 2026 March Madness Bracket</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c3892cc12dd7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Markowitz]]></title>
            <link>https://medium.com/@qfaclub/beyond-markowitz-b61d9128d704?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/b61d9128d704</guid>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[theory]]></category>
            <category><![CDATA[portfolio]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[markowitz]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Fri, 06 Mar 2026 13:49:53 GMT</pubDate>
            <atom:updated>2026-03-06T13:49:53.457Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This article was co-authored by </em><a href="https://www.linkedin.com/in/pinaki-ghosh-4a170221/"><em>Pinaki Ghosh</em></a><em> and </em><a href="https://www.linkedin.com/in/isaac-vergara/"><em>Isaac Vergara</em></a><em>.</em></p><p><strong>This is a continuation of our previous article on Markowitz Optimization Theory. To view that article, click here: </strong><a href="https://medium.com/@qfaclub/portfolio-optimization-series-part-1-the-markowitz-mean-variance-framework-applied-to-s-p-500-66ebdde22029"><strong>Portfolio Optimization Series — Part 1: The Markowitz Mean-Variance Framework Applied to S&amp;P 500 Sectors | by Quantitative Finance &amp; Analytics Club | Medium</strong></a></p><p><strong>I. What Markowitz Left Unanswered</strong></p><p>In Part 1, we applied mean-variance optimization to five S&amp;P 500 universes and found that the math works exactly as advertised. Given a set of stocks and a sample period, Markowitz finds the portfolio with the highest historical Sharpe ratio on the efficient frontier.</p><p>The healthcare portfolio illustrates what mean-variance optimization actually produces in practice. Out of the available stocks, the Max Sharpe portfolio collapsed into two names: AbbVie (ABBV) at 71.9% and Boston Scientific (BSX) at 27.6%, with Agilent Technologies (A) receiving a token 0.6%. The remaining stocks — including names like Amgen (AMGN), Baxter International (BAX), and Biogen (BIIB) — received zero allocation. The optimizer found that ABBV and BSX delivered the best risk-adjusted returns over the sample period and had no reason to look elsewhere. This is not diversification across healthcare. It is a concentrated bet on two companies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/854/1*LjgxzpHKLtrEWOts89YRqg.png" /><figcaption>Figure 1: Max Sharpe ratio weights for a Healthcare portfolio</figcaption></figure><p>The technology portfolio tells a similar story. Out of twelve available stocks, the Max Sharpe portfolio concentrated in just four: Broadcom (AVGO) at 43.1%, Amphenol (APH) at 32.2%, Alphabet (GOOG) at 18.1%, and Apple (AAPL) at 6.7%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/782/1*9cJSJ4eFYxX-KGVZCBLGhw.png" /><figcaption>Figure 2: Max Sharpe ratio weights for a Technology portfolio</figcaption></figure><p>This is not a flaw in the implementation; it is the expected behavior of an optimizer that sees only one dimension of risk: variance. When a stock has dominated a sector’s risk-adjusted returns over the sample period, the model has no reason to hold back. It bets the portfolio on that historical signal.</p><p>But why did those stocks win? Once we understand the sources of return, we can build portfolios that are more balanced without sacrificing much performance. That understanding runs through the Fama-French three-factor model.</p><p><strong>II. Fama-French Three-Factor Model</strong></p><p>In 1992, Eugene Fama and Kenneth French published findings that permanently changed how academics and practitioners think about equity returns. Building on the Capital Asset Pricing Model (CAPM), they showed that two simple stock characteristics — company size and book-to-market ratio — explained a large share of cross-sectional return variation that the single market factor left unaccounted for.</p><p>Their model expresses a portfolio’s excess return as a linear combination of three systematic factors:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4ykSURvI-XXwjuMbUS9RbQ.png" /><figcaption>Figure 3. Fama-French Three-Factor Model.</figcaption></figure><p>The central insight is that a portfolio’s return is not a single undifferentiated number. It is the sum of compensated exposures to systematic risks. A tech-heavy portfolio that rose 40% over five years may have done well mostly because tech stocks in general were booming, not because its specific stocks were uniquely well-chosen. A financial sector portfolio may be delivering a value premium that other, less concentrated instruments could replicate just as well.</p><p>For portfolio construction, this matters because factor exposures are, to a significant degree, interchangeable within a sector. If two stocks share nearly identical HML and SMB loadings, concentrating the entire portfolio in one of them delivers no factor diversification benefit that the other would not provide. The Markowitz optimizer cannot see this. Factor analysis can.</p><p><strong>III. Factor Fingerprints Across Sectors</strong></p><p>We ran FF3 regressions on equal-weight versions of each sector portfolio using daily returns from January 2020 through December 2024, with factor data sourced from <a href="https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/ftp/F-F_Research_Data_Factors_daily_CSV.zip">Kenneth French’s data library.</a> Running the regression on an equal-weight portfolio rather than the optimized one lets us characterize the sector’s underlying factor structure without the circular problem of using optimized weights to explain why the optimizer behaved the way it did.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nnm8kb4WbcZy_NnJKrPOyg.png" /><figcaption>Figure 4. FF3 Factor Loadings.</figcaption></figure><p>Technology’s beta HML of -0.28 confirms its growth character. The portfolio loads on stocks with low book-to-market ratios, and companies like Broadcom, Apple, and Alphabet, whose value lies in future earnings and intangible assets. Financials carry the strongest factor signal in the sample, with an HML beta of +0.61. Insurance companies, banks, and asset managers like the ones in our universe are textbook value stocks with high book-to-market ratios and mature cash flows. Healthcare sits almost exactly at zero on HML (-0.006), reflecting that the sector spans high-growth biotech and mature large pharma, and those exposures largely cancel each other out.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rEprCh9nkg3x1k8ALixF2g.png" /><figcaption>Figure 5. FF3 Regression Diagnostics by Sector.</figcaption></figure><p>Large Cap is the only portfolio with a statistically significant alpha at +5.5%. A 5-year sample is short, and we are cautious about reading too much into it, but the result is consistent with the idea that a cross-sector, quality-tilted, large-cap universe may capture return sources that the three factors do not fully explain.</p><p>Each sector has a distinct factor fingerprint, and that fingerprint explains a lot about why the optimizer behaves the way it does. Healthcare concentrates because its low R² means idiosyncratic stock performance, not systematic factors, drives returns; AbbVie delivered in that race over the sample period. Technology concentrates in Broadcom and Amphenol (APH) because those two names delivered the best growth-tilted, market-sensitive returns without the optimizer needing anything else.</p><p><strong>IV. What the Factors Reveal</strong></p><p>Once you understand what the optimizer is actually chasing, you can give it more informed constraints. The table below shows what happens when portfolios are rebuilt with factor targets derived from the regressions above and a 40% cap on any single stock.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KO4bTYKZhPrkN7rZMEZPOg.png" /><figcaption>Figure 6. Portfolio results after being rebuilt with factor targets.</figcaption></figure><p>The Sharpe cost of diversification is small across the board. Technology gives up 0.03, Large Cap gives up 0.01, and Diversified barely moves. The biggest drop is in Financials, where forcing weight away from Arthur J. Gallagher (AJG) has a real cost because that stock genuinely dominated the sample. But in every case, the maximum weight drops from something operationally problematic to something an actual portfolio manager could defend.</p><p><strong>V. Conclusion</strong></p><p>Factor analysis reframes the concentration problem. What looks like an optimizer making an aggressive bet on a single stock is, underneath, an optimizer efficiently acquiring a specific set of return exposures. The issue is that it found one vehicle for those exposures when several were available. Spreading weight across equivalent factor exposures costs almost nothing in Sharpe terms but produces a portfolio that is actually manageable in practice.</p><p>The numbers make the case quietly. Across five different universes, the Sharpe ratio barely moved while maximum single-stock weights dropped from the 50–80% range to 40%. That gap between the theoretical optimum and the practical portfolio is where factor awareness earns its keep.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b61d9128d704" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Systemic Shock: Simulating Credit Tail Risk with the Vasicek Model]]></title>
            <link>https://medium.com/@qfaclub/systemic-shock-simulating-credit-tail-risk-with-the-vasicek-model-35cfb054acc0?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/35cfb054acc0</guid>
            <category><![CDATA[credit-risk-modelling]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[monte-carlo]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[credit-cards]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Fri, 06 Feb 2026 12:51:21 GMT</pubDate>
            <atom:updated>2026-02-06T12:51:21.701Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This article was co-authored by </em><a href="https://www.linkedin.com/in/jordan-andrewww/"><em>Jordan Andrew</em></a><em>, </em><a href="https://www.linkedin.com/in/ammy-lin/"><em>Ammy Lin</em></a><em>, </em><a href="https://www.linkedin.com/in/farnoosh-memari/"><em>Farnoosh Memari</em></a><em>, and </em><a href="https://www.linkedin.com/in/tonantzinrealrojas/"><em>Tonantzin Real Rojas</em></a><em>.</em></p><p>Credit card portfolios represent one of the most significant exposure areas for retail banks, yet accurately predicting potential defaults remains a fundamental challenge. Traditional risk models often assume independent defaults, a simplified view that underestimates tail risk during economic crises and leaves financial institutions vulnerable to unexpected losses.</p><p>The 2008 financial crisis highlighted a critical flaw in this thinking: financial interconnectedness can trigger cascading defaults, a phenomenon that conventional probabilistic models fail to capture adequately. When one borrower fails to pay due to a systemic issue, their neighbor often does too.</p><p>To quantify this risk, we utilized a Vasicek-model Monte Carlo simulation. By analyzing credit card data, we explored a crucial question:<em> In the worst-case scenarios, how much capital should a bank reserve to cover losses?</em></p><h3>The Core Problem: Independent vs. Correlated Defaults</h3><p>In a stable economy, credit defaults are often viewed as isolated events. However, unlike traditional approaches that assume independent defaults, realistic stress testing must capture default dependency.</p><p>To model this, we utilized the Vasicek model, a framework that decomposes default risk into a systematic factor shared by all borrowers and an idiosyncratic component unique to each individual. Idiosyncratic and systemic risk are defined as:</p><ul><li>Idiosyncratic Risk: Risk specific to an individual borrower.</li><li>Systemic Risk: A common underlying economic factor that drives default clustering.</li></ul><p>This creates a more realistic approach where economic crises trigger simultaneous customer defaults.</p><h3>Methodology</h3><p>Using historical <a href="https://www.kaggle.com/datasets/uciml/default-of-credit-card-clients-dataset">credit card data</a> from clients in Taiwan (April–September 2005), we conducted a Monte Carlo simulation with 10,000 iterations to generate empirical loss distributions. The study stress-tested the portfolio against two primary levers:</p><ul><li>Probability of Default (PD): Ranging from 3% to 22.15%.</li><li>Asset Correlation: Ranging from 5% to 15%.</li></ul><p>The above values were determined from <a href="https://fred.stlouisfed.org/series/DRCCLACBS">literature</a>, and the 22.15% PD value was the average value found in our exploratory data analysis. While PD measures the likelihood of a single borrower not paying, correlation measures the extent to which individual asset values are driven by common systematic factors.</p><h3>Key Findings: The Multiplier Effect</h3><p>The results revealed a distinct difference between what drives “average” days and what drives “disaster” days.</p><ol><li>PD Drives Expected Loss: Probability of Default (PD) is the primary driver of the baseline average. As PD increases, the Expected Loss rises sharply; for example, at 5% correlation, increasing PD from 3% to 22.15% resulted in a more than seven-fold increase in expected losses. Put simply, if individual borrowers are riskier, the bank loses more money on a given day.</li><li>Correlation Drives Tail Risk: While the PD drives expected losses, correlation governs tail risk amplification. We can think of this like carrying eggs: High correlation is like putting all your eggs in one basket — if you trip (a systemic shock), you break them all at once. Low correlation means carrying them in separate baskets.</li><li>Tail Risk Amplification: When correlation rose from 5% to 15% at high default probabilities, our Value at Risk (VaR), which is the maximum expected loss, increased by up to 75%, and Conditional Value at Risk (CVaR) increased by up to 93%. This means that when borrowers’ financial behaviors are linked, banks must hold significantly more emergency cash to survive a crash.</li><li>The Gap Widens: In scenarios with low PD and high correlation, maximum losses exceeded expected losses by over 900%. This highlights how a “safe” looking portfolio can still suffer massive, unexpected losses if everyone stops paying at the same time.</li></ol><p>The analysis reveals that while the PD drives expected losses, correlation governs tail risk amplification. In other words, PD determines the baseline, but correlation determines the catastrophe.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e5AmNut4sAtonOsqUFMHyw.png" /><figcaption>Figure 1. Expected Loss by Combination. While the Probability of Default (PD) drastically increases the height of the loss bars, the correlation level ($\rho$) has almost no effect. This confirms that correlation does not change the <em>amount</em> of money a bank expects to lose on average, only the likelihood of those losses happening all at once.</figcaption></figure><h3>Visualizing this Risk</h3><p>The simulation demonstrated that as correlation increases, the “tail” of the loss distribution stretches. Box plots of the data revealed that increasing correlation consistently stretches the upper tail of losses, reflecting a higher chance of large, clustered loss events. This “fat tail” represents the catastrophic scenarios that standard models often miss.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/979/1*36RcRLKY5IIQOVqL-MzbMQ.png" /><figcaption>Figure 2. Loss Distributions by Combination. As correlation increases (moving left to right), the “whiskers” of the plot extend significantly, representing catastrophic outlier losses that average-based models miss.</figcaption></figure><h3>Measuring the Extremes: VaR and CVaR</h3><p>To quantify the danger, we used two specific metrics:</p><ul><li>Value at Risk (VaR): The maximum loss at a specified confidence level (e.g., 99%).</li><li>Conditional Value at Risk (CVaR): Also known as Expected Shortfall, this measures the average loss <em>beyond</em> the VaR threshold.</li></ul><p>At a 99% confidence level, a scenario with high PD (22.15%) and high correlation (15%) resulted in a VaR exceeding NT$ 2.8 billion, which was nearly seven times higher than the VaR of NT$ 414 million observed in the low-risk scenario (3% PD, 5% correlation). This confirms that ignoring correlation allows banks to drastically underestimate the capital buffers required for survival.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wP-qwoePDtkc1sdp2BZHmg.png" /><figcaption>Figure 3. VaR 99% by Parameter Combination. Unlike Expected Loss, VaR is highly sensitive to systemic factors. As we move from left to right (increasing correlation from 5% to 15%), the tail risk spikes significantly. In the worst-case scenario (purple bars), high correlation pushes the potential loss to over NT$2.8 billion.</figcaption></figure><h3>Implications for Financial Institutions</h3><p>The practical applications of this study extend directly to regulatory compliance and capital planning.</p><ul><li>Capital Allocation: Financial institutions require robust tools to meet Basel capital adequacy requirements. Our data shows that Maximum Loss is sensitive to the dependence structure generated by asset correlation.</li><li>Regulatory Compliance: The Vasicek model underlies the Basel internal ratings-based (IRB) approach, which uses this structure to determine capital requirements.</li><li>Stress Testing: The study provides a practical tool for illustrating how systemic factors amplify losses in retail credit portfolios under stress conditions.</li></ul><h3>Outlook</h3><p>While this simulation focused on correlation levels between 5% and 15%, real-world crises can drive correlations even higher. Future research could systematically vary correlation across a wider spectrum (e.g., 1% to 30%) to better understand the drivers of catastrophic, clustered-loss scenarios. Additionally, integrating network analysis could provide a more granular view of systemic risk, allowing researchers to map exactly how defaults propagate through interconnected economic links rather than relying solely on global correlation parameters.</p><p>Ultimately, this research underscores a critical lesson: Traditional risk models that assume independent defaults underestimate tail risk. As such, in a hyper-connected global economy, accurate capital planning depends on modeling the correlation between assets just as much as the assets themselves.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=35cfb054acc0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Portfolio Optimization Series — Part 1: The Markowitz Mean-Variance Framework Applied to S&P 500…]]></title>
            <link>https://medium.com/@qfaclub/portfolio-optimization-series-part-1-the-markowitz-mean-variance-framework-applied-to-s-p-500-66ebdde22029?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/66ebdde22029</guid>
            <category><![CDATA[markowitz]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[money]]></category>
            <category><![CDATA[portfolio]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Wed, 19 Nov 2025 18:54:32 GMT</pubDate>
            <atom:updated>2025-11-19T18:54:32.082Z</atom:updated>
            <content:encoded><![CDATA[<h3>Portfolio Optimization Series — Part 1: The Markowitz Mean-Variance Framework Applied to S&amp;P 500 Sectors</h3><p>This article was co-authored by <a href="https://www.linkedin.com/in/pinaki-ghosh-4a170221/">Pinaki Ghosh</a>, <a href="https://www.linkedin.com/in/gaurav-law/">Gaurav Law</a>, <a href="https://www.linkedin.com/in/ammy-lin/">Ammy Lin</a>, and <a href="https://www.linkedin.com/in/isaac-vergara/">Isaac Vergara</a>.</p><h3>I. Series Introduction: The Evolution of Portfolio Theory</h3><p>What started as intuition slowly changed to adapt to modern-day technologies. The first investment portfolios — based solely on gut feeling — shifted toward a more data-driven approach after 1950. Through the founding of investment trusts and mutual funds, several newer approaches have been taken to change the idea of modern portfolio construction.</p><p>Led by American economist Harry Markowitz, the first idea of modern investment portfolios was to <a href="https://www.sciencedirect.com/science/article/pii/S0378426697000484">maximize return by minimizing variance</a>. Markowitz’s Modern Portfolio Theory (MPT) is still being developed through other methods; Eugene Fama and Kenneth French expanded on the Capital Asset Pricing Model (CAPM), attempting to show a residual learning process for returns across periods. These theories continue to be expanded today.</p><p>As students, practitioners, and quantitative analysts, we want to help optimize a portfolio based on Markowitz’s original theory. Below, you will find a detailed explanation of the history, concepts, and roadmap of modern-day portfolio construction.</p><h4>The Pre-Markowitz Era and the 1952 Revolution</h4><p>Before the introduction of MPT in 1952, investment strategies were largely intuitive and lacked a rigorous mathematical foundation. The general idea was to select low-risk, high-return stocks, often guided by personal judgment and anecdotal evidence rather than rigorous analysis. Investors relied on the adage, “don’t put all your eggs in one basket,” which emphasized the importance of diversification but did not provide a systematic approach to achieving it.</p><p>As a result, many investors concentrated their portfolios in a few assets, believing that selecting a handful of promising stocks would yield the best returns. However, this approach often led to significant risk exposure, as individual stock performance could be volatile and unpredictable; little consideration was given to how different investments interacted.</p><p>Markowitz’s theory transformed investment strategy by introducing a mathematical framework that quantified the risk-return tradeoff, the principle that higher returns typically come with higher risk. He demonstrated that diversification, guided by an understanding of how asset returns move in relation to one another, could reduce overall portfolio volatility without sacrificing expected returns. This approach enabled investors to construct more resilient portfolios, marking a shift from intuition-based decision-making to a systematic, data-driven methodology.</p><p>Building on this foundation, Markowitz also introduced the concept of the efficient frontier, a graphical representation of all possible portfolios that offer the best expected return for a given level of risk or the lowest risk for a given level of return. Portfolios that lie on the efficient frontier are considered optimal because they maximize return while minimizing unnecessary risk.</p><p>In contrast, portfolios below the frontier are inefficient, as they either expose investors to more risk than necessary or yield lower returns than could be achieved with a more optimal asset combination. By plotting various combinations of assets based on their expected returns, variances, and correlations, the efficient frontier not only helps investors clearly visualize the trade-offs between risk and return but also serves as a practical tool for portfolio optimization, reinforcing the broader shift toward systematic, evidence-based investment decision-making.</p><h4>Core Concepts of Markowitz Theory</h4><p>Portfolio risk isn’t additive — this non-linearity creates the entire opportunity for optimization. When Asset A drops 2% and Asset B rises 1.5%, they partially offset. This only happens when the correlation is less than 1. The mathematical representation captures all pairwise interactions:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/508/1*ctqJib1KmWbKJa3OyNRkLQ.png" /></figure><p>This double summation — every asset with every other asset, including itself — is why portfolio math gets computationally intensive. With 500 stocks (like the S&amp;P 500), you are calculating 250,000 covariance terms.</p><p>Selecting variance as risk was not arbitrary. Markowitz needed something that both made optimization solvable with 1950s computing power and captured uncertainty in a single number.</p><p>Standard deviation (variance’s square root) has interpretable units — it’s in the same scale as returns. A 20% standard deviation means roughly 68% of returns fall within 20% of the mean (assuming normality). This interpretability made it practical for portfolio managers who needed to communicate risk to clients.</p><h4>The Optimization Problem</h4><p>Two equivalent approaches exist for finding efficient portfolios:</p><p><strong>First approach:</strong> Set a maximum risk level, then find the highest return possible within that constraint. Hedge funds often work this way — they promise investors they won’t exceed 15% volatility, then maximize returns within that boundary.</p><p><strong>Second approach:</strong> Set a target return, then find the least risky way to achieve it. Pension funds typically use this — they need 7% returns to meet future obligations, so they find the safest portfolio that delivers it.</p><p>Both approaches trace out the same efficient frontier, just from different starting points. The choice between them isn’t about mathematical superiority — it’s about which question you’re actually trying to answer.</p><p>The Sharpe Ratio (SR), introduced by William Sharpe in 1966, became finance’s standard measure of risk-adjusted returns:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/324/1*f0A_P0nbL_3J1VyeHvNtjQ.png" /></figure><p>Where:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/666/1*-yDHoU0AtSFuRDf-oTfTuA.png" /></figure><p>How much excess return per unit of risk? An SR of 1.0 means 1% extra return for every 1% volatility. The key property is scale independence:</p><ul><li>Portfolio A: 10% return, 10% volatility with 2% risk-free rate → SR = 0.8</li><li>Portfolio B: 20% return, 20% volatility with 2% risk-free rate → SR = 0.8</li></ul><p>Portfolio B isn’t better — it’s just A with leverage. Sharpe Ratio correctly shows the same efficiency.</p><p>The <strong>maximum Sharpe portfolio</strong> sits where a line from the risk-free rate touches the frontier. This ‘tangency portfolio’ offers unique flexibility.</p><h4>The Efficient Frontier</h4><p>The frontier’s hyperbolic shape encodes a fundamental law: Risk reduction has diminishing returns. Moving from one stock to two reduces risk dramatically. Moving from 50 to 51 stocks barely moves the needle.</p><p>Mathematically, the frontier satisfies:</p><ul><li>No portfolio northwest with more return and less risk exists</li><li>Every frontier portfolio minimizes variance for its return level</li><li>The frontier is convex</li></ul><p>Below-frontier portfolios are irrational under mean-variance preferences. Holding them means accepting unnecessary risk or foregoing returns. However, most real portfolios sit below the frontier — due to constraints, transaction costs, or behavioral biases.</p><h3>II. Application to S&amp;P 500 Sectors: Empirical Results</h3><p>We applied the Markowitz mean-variance optimization framework to a diversified basket of 22 stocks selected systematically across all S&amp;P 500 sectors. This analysis demonstrates how systematic portfolio construction improves risk-adjusted returns compared to naive allocation strategies.</p><h4>Portfolio Construction Methodology</h4><p>The diversified basket was constructed using a <strong>sector-balanced approach</strong>:</p><ol><li>Identified all 11 Global Industry Classification sectors in the S&amp;P 500</li><li>Selected top stocks from each sector proportionally (2 stocks per sector)</li><li>Total portfolio: 22 stocks representing broad market exposure</li></ol><p>This methodology provides sector diversification, follows a systematic approach, and delivers representative market exposure.</p><h4>Analysis of the Efficient Frontier</h4><p><strong>The efficient frontier </strong>shows the trade-off between risk (volatility on the x-axis) and expected return (y-axis).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A_isdo00R0z2ECAj4ehyGg.png" /><figcaption>Figure 1: Efficient Frontier for a Diversified Basket (2 stocks from each S&amp;P sector)</figcaption></figure><p>Why the curve beats individual dots: stocks don’t move together perfectly. By combining uncorrelated or less-correlated names (e.g., a defensive telecom like Verizon (VZ) with a growth name like Amazon (AMZN) or a pharma like AbbVie (ABBV)), a portfolio can reduce risk without proportionally cutting return.</p><h4>Key Observations on the Efficient Frontier:</h4><ol><li><strong>Diversification benefit is strong at low risk.<br></strong>The curve rises steeply on the left: Moving from the green star to slightly more risk adds meaningful expected return. That’s the ‘free’ improvement you get from mixing unlike businesses (healthcare, staples, tech, energy, telecom).</li><li><strong>Several single stocks look inefficient compared to the frontier.<br></strong>Names like APA (APA) and Albemarle (ALB) show high expected returns but need very high volatility to get there. The blue curve sits above and to the left of these points, meaning a diversified portfolio can target similar returns with much less risk.</li><li><strong>Why AbbVie keeps showing up in optimal mixes.<br></strong>AbbVie (ABBV) has a high expected return at mid-range volatility, so it sits near the top-left cluster of stocks — closer to the curve than many others. That makes it a natural building block for the optimizer.</li></ol><ul><li><strong>Roles of low-beta names.<br></strong>Verizon (VZ) has a low expected return but low volatility. Stocks like Verizon are helpful for the Min Vol portfolio: They pull portfolio risk down without destroying the overall return once combined with higher-return names.</li></ul><h4>What does it all mean?</h4><p>This chart demonstrates <em>what</em> the efficient frontier is, <em>why</em> diversification works, and <em>how</em> a diversified portfolio here can target 22.43<strong>%</strong> expected return at 20.03<strong>%</strong> volatility — superior to many single-stock choices at similar or higher risk.</p><h4>Market Cap Weighted Allocation</h4><p>Mega-cap names dominate a market capitalization-weighted portfolio.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8JBk4k8qy-1HHqn4acEtKg.png" /><figcaption>Figure 2: Market Cap Weights for a Diversified Basket</figcaption></figure><p>This donut chart shows how a market capitalization-weighted diversified portfolio would typically distribute investments across large S&amp;P 500 constituents before optimization. It helps establish a baseline: how much each stock contributes to the overall portfolio simply based on its size in the market, not on its risk-return efficiency. Amazon (AMZN) dominates with a 52.1% weight, reflecting its large market value.</p><h4>Observations</h4><ol><li><strong>High Concentration in Mega-Caps<br></strong>Over half the portfolio is concentrated in Amazon, showing that a market-cap-weighted strategy naturally favors mega-cap growth firms. This is typical in diversified indices but creates a heavy tilt toward a few dominant names.</li><li><strong>Sets Optimization Baseline<br></strong>When compared later to the efficient frontier portfolios, we can see how optimization reallocates weights — away from pure market size and toward risk-adjusted return efficiency.</li><li><strong>Sector Skew<br></strong>Because of Amazon’s dominance and the inclusion of AbbVie and Chevron, the portfolio is overweighted in technology, healthcare, and energy<strong>,</strong> even before optimization. Defensive or cyclical sectors, like industrials or utilities, barely register in percentage terms.</li><li><strong>Implied Diversification Gap<br></strong>While there are more than 20 holdings, the effective diversification is much lower because of concentration. In practice, the top three holdings (Amazon, AbbVie, Chevron) account for nearly 70% of the total exposure.</li></ol><h4>Optimized Portfolios Deep-Dive</h4><p>When constructing an optimized portfolio, the choice of optimization objective fundamentally reshapes asset allocation. Rather than simply weighting assets by market capitalization, modern portfolio theory allows us to target specific goals: maximizing risk-adjusted returns through the Sharpe ratio, or minimizing portfolio volatility for capital preservation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v-7wD2RTr3B40hYK3am8yg.png" /><figcaption>Figure 3: Optimized Portfolios for a Diversified Basket (Min Vol perspective).</figcaption></figure><p>The two charts illustrate how optimization reshapes portfolio composition once we move away from simple market-cap weighting and focus instead on risk-adjusted efficiency (for the Max Sharpe case) or stability (for the Min Volatility case). Together, they highlight how the Markowitz optimization framework reallocates capital toward the most effective risk-return combinations.</p><h4>1. Max Sharpe Portfolio</h4><p>In the Max Sharpe portfolio, weights are assigned to maximize the Sharpe ratio — the return earned per unit of risk. This portfolio aims for the most efficient point on the efficient frontier, represented by the red star in the earlier chart.</p><p>AbbVie carries the highest weight at 50.6%, indicating that its combination of high return and moderate volatility makes it the strongest individual contributor to portfolio efficiency. Furthermore, when paired with Altria (MO), the stocks deliver the best combination of expected return and volatility.</p><p>However, in practice, dedicating 70% of the total portfolio to just two stocks exposes one to high risk.</p><h4>2. Min Volatility Portfolio</h4><p>The Min Volatility portfolio takes the opposite approach: It minimizes total risk regardless of return. The objective here is capital preservation with smoother performance.</p><p>Unlike the Max Sharpe version, which was dominated by a few high-return names, this portfolio spreads exposure more evenly across low-beta, dividend-oriented companies. The result is a portfolio that minimizes swings in value even if its expected return is lower. For example, Verizon (VZ) dominates with 30.7%, reflecting its consistent defensive returns and low price fluctuations.</p><p>This portfolio prioritizes consistency and downside protection over aggressive growth, as that of the Max Sharpe Portfolio.</p><h4>3. Comparison between the Optimized Portfolios</h4><p>A tabular comparison of the two optimized portfolios across different characteristics:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wPPe7ZP6WxM_hqUdoyKlHg.png" /></figure><p>Together, these two optimized portfolios demonstrate the core principle of the Markowitz framework:</p><ul><li>The Max Sharpe portfolio seeks efficiency by concentrating on the best-performing names.</li><li>The Min Volatility portfolio spreads exposure across defensive assets to lower overall variance.</li></ul><p>When compared to the Market Cap Weighted portfolio (dominated by Amazon), both optimized portfolios show how strategic allocation — based on risk and return metrics rather than size — can transform performance outcomes.</p><p>This contrast reinforces that optimization creates intentional, data-driven portfolios rather than passive replications of the index.</p><h3>III. Beyond Markowitz</h3><h4>Lessons from the Diversified Portfolio</h4><p>The diversified case encapsulates the central insight of modern portfolio theory — that<strong> </strong>risk and return are not properties of individual assets, but of how those assets interact<strong>.</strong> The efficient frontier, when combined with optimized allocations, reveals that diversification is not simply about owning more stocks, but about owning the <em>right</em> mix of them.</p><p>The progression from Market Cap → Max Sharpe → Min Volatility can thus be seen as a journey across the efficient frontier:</p><ul><li>The market-cap mix sits <em>below the curve</em> — suboptimal due to unmanaged concentration risk.</li><li>The Max Sharpe mix lies <em>on the upper middle of the curve</em> — high return per unit of risk.</li><li>The Min Volatility mix sits <em>at the lower-left end</em> — minimal risk, modest return.</li></ul><p>The Market Cap Weighted portfolio demonstrates that size-based exposure, while representative of the market, can lead to structural concentration and inefficiency. The Max Sharpe Portfolio proves that strategic reallocation toward assets like AbbVie and Altria<strong> </strong>— which deliver the highest returns per unit of risk — can substantially enhance performance. Meanwhile, the Min Volatility Portfolio highlights the defensive potential of low-beta names such as Verizon and Abbott Laboratories, showing that stability and diversification can coexist even with modest returns.</p><p>Together, these results illustrate the risk-reduction power of thoughtful diversification and the performance gains from optimization. The shift from a cap-weighted structure to mathematically optimized allocations produces portfolios that sit directly <em>on</em> the efficient frontier, achieving the best trade-offs achievable within the data.</p><p>This diversified case, therefore, sets the benchmark for the subsequent sector analyses — technology, healthcare, financials, and large-cap — where we examine how the same optimization principles apply within specific market segments and how each sector’s risk-return profile shapes its unique frontier.</p><h4>The Gap Between Theory and Practice</h4><p>While the diversified and optimized portfolios demonstrate the power of mean–variance optimization in theory, practical implementation often encounters several limitations. Real-world investors face estimation errors, transaction costs, and changing market conditions that make it difficult to maintain the precise weightings implied by the mathematical model.</p><p>The optimizer assumes that expected returns, volatilities, and correlations are known and stable — but in reality, they are only estimates based on historical data, and even small errors in these inputs can lead to disproportionately large shifts in the optimal weights. This is why theoretical portfolios, like the Max Sharpe mix, often appear overly concentrated or unstable when applied to live markets.</p><p>Our diversified S&amp;P 500 portfolio revealed Markowitz’s fundamental challenge: Even with stocks across eleven sectors, the optimizer concentrated 50.6% in AbbVie and 18.7% in Altria. Two stocks captured 69.3% of a supposedly diversified portfolio. This concentration reflects the optimizer finding the highest historical Sharpe ratios and betting heavily on them, regardless of sector representation.</p><p>In Part 2, we’ll examine how this concentration pattern unfolds across individual S&amp;P sectors — tech, healthcare, and finance — and introduce Fama-French factors to explain why certain stocks dominate these optimizations. Their multi-factor framework — capturing size, value, and profitability — identifies return drivers that variance alone overlooks. We’ll show how factor models yield more balanced, implementable portfolios while preserving the mathematical rigor of Markowitz’s approach.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=66ebdde22029" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Unearthing Alternative Data: How Alt Data Is Redefining What We Know]]></title>
            <link>https://medium.com/@qfaclub/unearthing-alternative-data-how-alt-data-is-redefining-what-we-know-c7850cdfa8d8?source=rss-3d52e7950e8b------2</link>
            <guid isPermaLink="false">https://medium.com/p/c7850cdfa8d8</guid>
            <category><![CDATA[alternative-data]]></category>
            <dc:creator><![CDATA[Quantitative Finance & Analytics Club]]></dc:creator>
            <pubDate>Tue, 04 Nov 2025 02:06:58 GMT</pubDate>
            <atom:updated>2025-11-04T15:03:53.171Z</atom:updated>
            <content:encoded><![CDATA[<p>Link to podcast: <a href="https://open.spotify.com/episode/1eGjSaWoVMEFYfEHPUm3Qx?si=b8lf52uGS4OZ2KxnWLH1eg">https://open.spotify.com/episode/1eGjSaWoVMEFYfEHPUm3Qx?si=b8lf52uGS4OZ2KxnWLH1eg</a></p><h3>From Walmart Parking Lots to Satellites in Space</h3><p>On the first day of our “Alternative Data” class at Duke, we were asked a deceptively simple question: <em>What is alternative data?</em></p><p>Back then, our best answer was something like this: <em>It’s when you count the number of cars in a Walmart parking lot using satellite images and then trade Walmart’s stock before earnings are released.</em></p><p>A neat trick, right? But as the semester unfolded, we realized alternative data — <em>alt data</em> — is much more than just a clever hedge-fund story. It’s not just about counting cars. It’s about measuring the world itself, in all its messy, complex, and data-rich forms.</p><h3>When Wall Street Looks Down from Space</h3><p>Before this class, Arko had already spent time in an asset-management alt-data team, scraping open-source information, structuring feeds, and trying to make sense of the unstructured.</p><p>“Satellite imagery,” he says, “is the glamorous face of alt data. It’s what gets people’s attention.”</p><p>Indeed, hedge funds now subscribe to private satellite feeds to monitor everything from factory activity in China to port congestion in Russia. Some firms use optical sensors to count oil tankers leaving Russian ports, a method that helps them estimate the “ghost ships” of global trade, carrying sanctioned oil through opaque networks.</p><p>Danish, who worked in commodities trading, recalls using satellite data to measure supply in real time. “We were literally tracking the flow of oil using images from orbit,” he says. “When you can see tankers moving, or disappearing, you can anticipate market movements days before the data becomes public.”</p><h3>Beyond Images: Satellites That Sense</h3><p>But not all satellites just take pictures. Many are equipped with <strong>spectral sensors</strong> that detect things invisible to the human eye, like methane leaks or particulate emissions.</p><p>As Arko explains, “Some companies now use post-processing models on these satellite captures to estimate methane levels over industrial zones. They look at cloud opacity, light intensity, and surface reflection to build a proxy for factory emissions. That’s how you get an edge on real-time industrial activity.”</p><p>What was once science fiction, reading the Earth’s pulse from orbit, is now a data product that can move billions of dollars.</p><h3>The Ethics of Watching Everything</h3><p>But the conversation soon turned from innovation to unease.</p><p>Danish notes that “this kind of monitoring doesn’t just reveal company performance , it can expose community behavior, migration patterns, or even military movements.”</p><p>That’s where privacy concerns start to blur the line between <em>observation</em> and <em>surveillance</em>.</p><p>“Imagine,” Vishesh adds, “that parking your car in a lot might nudge a hedge fund’s algorithm to buy or sell Walmart stock. You never agreed to be part of that feedback loop.”</p><p>And while the data is technically public, access isn’t equal. A few large funds pay hundreds of thousands of dollars for these feeds, giving them a 3–5% market edge purely from data asymmetry, not from analytical skill. “It’s like insider trading for the data-rich,” says Vishesh. “Not illegal, but inaccessible.”</p><h3>Who Regulates the Skies?</h3><p>The group raises a bigger question: <em>Who decides what’s fair game in the world of alt data?</em></p><p>Should regulators like the SEC maintain a central registry of the alternative datasets firms use for trading? That could increase transparency without banning innovation.</p><p>But even if such oversight existed, the inherent data gap would remain. Retail investors can’t buy daily methane readings or 10-meter resolution port imagery.</p><p>“The sky might be public,” Vishesh says, “but the pixels are priced.”</p><h3>From Markets to Humanitarian Impact</h3><p>Not all satellite data needs to serve hedge funds. The same feeds could help NGOs track disaster damage, floods, or wildfires in real time, optimizing aid distribution.</p><p>“Alternative data could be <em>alternatively used</em>,” Arko remarks. “Imagine if the UN or disaster-relief agencies had access to these feeds for free. We could optimize food distribution after a flood the same way hedge funds optimize trades.”</p><p>But democratizing access raises another paradox: the more open the imagery, the greater the privacy risk. “Nobody walks outside thinking a satellite might be watching,” Danish points out. “We deserve some degree of anonymity.”</p><h3>Personal Projects: Turning Learning into Action</h3><p>By the end of the course, each of us pursued a project exploring a different dimension of alternative data:</p><ul><li><strong>Vishesh </strong>and his partner<strong> </strong>analyzed how <strong>political leanings and crime rates</strong> influence real-estate prices, building models that re-weight these social variables to predict investment potential.</li><li><strong>Danish</strong> used <strong>in-store surveillance footage</strong> to map customer movement and optimize product placement, transforming behavioral patterns into marketing insights.</li><li><strong>Arko</strong> combined <strong>satellite light pollution</strong> and <strong>PM2.5 air-quality data</strong> to create proxy signals for factory output, uncovering positive correlations between environmental indicators and production levels.</li></ul><p>Three very different problems, one unifying theme: using unconventional data to see what traditional data cannot.</p><h3>From Curiosity to Consciousness</h3><p>When we began in August, we thought alternative data was just about alpha generation. But as our discussions deepened, the field revealed itself as something far more complex, part science, part philosophy.</p><p>It’s about how far we’re willing to go in quantifying the world around us. It’s about the trade-off between <em>insight</em> and <em>intrusion</em>.</p><p>As Arko puts it: “If you don’t invent new technologies, you won’t have new problems , but you also won’t have new opportunities to make life better.”</p><h3>Final Thoughts: The Data-Rich Sky Above Us</h3><p>Alternative data blurs the line between finance, ethics, and society. It’s a mirror that reflects both human curiosity and human consequence.</p><p>And as we wrapped our semester, and this conversation, we couldn’t help but wonder:</p><blockquote><em>The sky was once a symbol of freedom. Now, it’s our largest data feed.<br>Is that progress, or a new form of dependence?</em></blockquote><p>Perhaps it’s both.</p><p><em>“Data is no longer just numbers. It’s everything else now.”</em><br> — <em>The All Data Podcast Team <br>(Arko Bhattacharya, Vishesh Gupta, Danish Maknojia)</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c7850cdfa8d8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>