<?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 J.R. Willett on Medium]]></title>
        <description><![CDATA[Stories by J.R. Willett on Medium]]></description>
        <link>https://medium.com/@jr.willett?source=rss-1961ee6f2887------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*LTQEi3W1M0KWgMN8.</url>
            <title>Stories by J.R. Willett on Medium</title>
            <link>https://medium.com/@jr.willett?source=rss-1961ee6f2887------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 15:50:42 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jr.willett/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[AI Robot Jesus will save us from Hyperinflation]]></title>
            <link>https://medium.com/@jr.willett/ai-robot-jesus-will-save-us-from-hyperinflation-1ca9b81541a4?source=rss-1961ee6f2887------2</link>
            <guid isPermaLink="false">https://medium.com/p/1ca9b81541a4</guid>
            <category><![CDATA[predictions]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[inflation]]></category>
            <category><![CDATA[deflation]]></category>
            <dc:creator><![CDATA[J.R. Willett]]></dc:creator>
            <pubDate>Mon, 11 Dec 2023 17:58:01 GMT</pubDate>
            <atom:updated>2023-12-11T17:58:01.444Z</atom:updated>
            <content:encoded><![CDATA[<p>Careless fiscal policy has set us up for massive inflation, but productivity growth from AI could bring massive deflation, potentially balancing out our societal mistakes.</p><figure><img alt="A Robot Messiah, an AI Jesus, stands with arms outstreched, ready to bring enough deflation to heal our inflation sins (generated by Midjourney)" src="https://cdn-images-1.medium.com/max/1024/1*iIc10H40TGqdtlCZBJXbwQ.jpeg" /><figcaption>A Robot Messiah, an AI Jesus, stands with arms outstretched, ready to bring enough deflation to heal our inflation sins (generated by Midjourney)</figcaption></figure><p>I have long expected our current era of the US <a href="https://www.google.com/search?q=exporting+inflation">exporting inflation</a> to end in tears. It appears to me that the US can only run absurd deficits based on two main pillars:</p><ol><li>The whole world wants dollars for trade</li><li>Productivity growth means it is easier to pay back debts as our GDP gets bigger</li></ol><p>Nothing lasts forever, and the eventual end of dollar-denominated trade would presumably be a disaster for the US government, which will have to re-import all the dollars (and inflation) it previously exported. Big debts would continue to come due but the appetite for new US debt would be greatly reduced, causing any number of painful desperate measures, possibly including:</p><ul><li>Heavy taxation</li><li>Severe currency debasement (hyperinflation)</li><li>Seizing of assets (think <a href="https://www.google.com/search?q=US+gold+confiscation">1930s gold confiscation</a>, or when <a href="https://www.google.com/search?q=Cyprus+seized+bank+account+balances">Cyprus seized bank account balances</a>).</li><li><a href="https://www.google.com/search?q=changing+cpi+methodology+hides+real+inflation">Changing how inflation is calculated</a> (the Consumer Price Index) to make inflation look smaller (and reduce inflation-indexed government expenditures like TIPS).</li><li>Capital controls (Limits on movements of currency over borders)</li><li>Trade controls (<a href="https://www.gresham.ac.uk/watch-now/when-currency-empires-fall">See Britain’s “Sterling Area” as they faced a similar situation as the Pound Sterling was used less and less in global trade</a>)</li><li>Price controls (Government sets prices to try to stop inflation, <a href="https://www.google.com/search?q=price+controls+create+black+market">creating a black market</a>)</li></ul><p>The US has done all of these things to varying degrees in the past, and we would expect all of this and more as the world is “de-dollarized”, especially given the huge deficits run by the US government.</p><h3>Our Coming Fiscal AI Messiah</h3><figure><img alt="Midjourney imagines a robot messiah. Basically creepy Jesus with lots of wires." src="https://cdn-images-1.medium.com/max/1024/1*sImpK1rwPktulS4fuazx3A.jpeg" /><figcaption>Midjourney imagines a robot messiah. Basically creepy Jesus with lots of wires.</figcaption></figure><p>HOWEVER, the use of dollars in global trade is only one of the two “pillars” mentioned above allowing the US to run up huge debts. The other pillar is the growth of productivity. What if that pillar were to get a LOT bigger?</p><p>Productivity, if it grows fast enough, actually <a href="https://www.google.com/search?q=productivity+causes+deflation">causes deflation</a>! And the AI boom could bring the <a href="https://www.google.com/search?q=projected+impact+of+AI+on+productivity">mother of all productivity booms</a>. Furthermore, it is my personal opinion that despite all the hype, AI actually remains vastly under-hyped. I think the existing projections <a href="https://www.google.com/search?q=humans+fail+to+think+exponentially">fail to think exponentially</a>, and over the course of the coming years, whether or not we get “<a href="https://www.google.com/search?q=%22runaway+ai%22">runaway AI</a>”, we will probably have something like “runaway productivity” which implies runaway deflation.</p><p>Essentially, I think the historical collapse of the US Dollar (hyperinflation) could coincide with with the exponential productivity growth from AI (hyperdeflation), essentially balancing each other out.</p><p>As far as I can tell, only a few outlier prognosticators are predicting each of these extreme events, and I don’t see ANYBODY predicting both at the same time. However, I think the possibility is real, and frankly I want the credit of having been the first to predict it, in case it actually comes to pass.</p><h3>How to position yourself</h3><p>So let’s say you believe in both the possibility of hyperinflation from the dollar collapse AND the possibility of hyperdeflation from the AI boom. How does one best position for this outcome? Here’s how I think about it (but please note I’m not an investment advisor, and I’m not anywhere near the mainstream in my views):</p><ul><li>Buy whatever you think will replace the dollar. I think cryptocurrency has a shot, and personally I think Ethereum is well-positioned (but please note crypto is volatile and risky!)</li><li>Buy real estate. It’s a traditional inflation hedge, and it should do well in a dollar collapse scenario. However I personally think US real estate is currently overvalued and could crash big. If it does, I think buying the crash could be a great long-term investment</li><li>Buy companies at the forefront of AI. Lots of people are already doing this, and I think there could be a crash (like the dotcom crash) before it really takes off. Buying that crash could be a fabulous long-term investment.</li><li>Buy precious metals. Whatever happens, they probably won’t hyperinflate.</li><li>Basically, buy anything that isn’t the US dollar. However, even if you just have US dollars, you might be fine. Runaway productivity could give us crazy socialist dreams like universal basic income, and work could become completely optional for many people.</li></ul><p>I believe AI could give us the kind of living standards that used to be only achievable by the upper crust of societies that engaged in human slavery on a massive scale. I expect AI will be trained to effectively be slaves for humanity, and to be content in that role, and I expect our fiscal sins to be forgiven in the process. Thank you AI Robot Jesus.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1ca9b81541a4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Timing This Crypto Bubble (February 2021)]]></title>
            <link>https://medium.com/@jr.willett/timing-this-crypto-bubble-february-2021-ab443362bf76?source=rss-1961ee6f2887------2</link>
            <guid isPermaLink="false">https://medium.com/p/ab443362bf76</guid>
            <category><![CDATA[cryptocurrency]]></category>
            <category><![CDATA[ethereum]]></category>
            <dc:creator><![CDATA[J.R. Willett]]></dc:creator>
            <pubDate>Tue, 09 Feb 2021 07:47:45 GMT</pubDate>
            <atom:updated>2021-02-09T09:08:17.051Z</atom:updated>
            <content:encoded><![CDATA[<p>I don’t tend to make a lot of predictions. Every prediction is a new opportunity to look stupid, and I’ve been wrong plenty of times.</p><p>However, in August 2018, I announced very specific prices which I believed would be close enough to the bottom for me, personally, to buy Ethereum: <a href="https://medium.com/@jr.willett/time-for-me-to-buy-cryptocurrencies-august-2018-1d31cc1ecb8d">https://medium.com/@jr.willett/time-for-me-to-buy-cryptocurrencies-august-2018-1d31cc1ecb8d</a></p><figure><img alt="ETH bottom" src="https://cdn-images-1.medium.com/max/962/1*vMwYp36mL6X3sD422JGGHA.png" /></figure><p>In hindsight, my plan looks downright prescient. Life circumstances prevented me from executing as I intended, but anybody who followed this plan made out like a bandit. Partly I just got lucky, but partly I was simply following a pattern in cryptocurrency bubbles which seems to keep happening.</p><p>The pattern is very simple: a HUGE run-up of 2 orders of magnitude or more (100x or more), followed by a massive crash of about an order of magnitude (10x crash). That’s it. It has happened three times so far in the history of cryptocurrency, and I believe we are now entering the fourth bubble. <strong>DISCLAIMER: These magnitudes are extremely fuzzy. I have seen 1000x gains, and I’ve personally held cryptocurrencies that fell 100x!</strong></p><p>So, how do we try to exploit the current bubble? Here are my thoughts:</p><p>I’ll start by saying that I think <strong>Ethereum is the right bet</strong> if you are going to buy just one cryptocurrency. When I make a list of the most innovative cryptocurrency projects, almost all of them are building on Ethereum. I love bitcoin, and it gets a lot of press, but <strong>Ethereum is where the innovation is happening</strong>. I don’t know how soon it will happen, but I believe Ethereum deserves to eventually surpass bitcoin in market cap (known as “The Flippening” which very nearly happened back in 2017). If I am right, Ethereum deserves to rise at least 4x more than bitcoin in the long run. <strong>DISCLAIMER: I am not unbiased. I am long Ethereum.</strong></p><p>IF our historical pattern holds (past performance does not guarantee future results!), we’d expect this bubble to peak at more than 100x over the most recent Ethereum post-bubble low ($95). That would suggest a price target of at least $9,500 (as I write this the price is $1755).</p><p>Even knowing gains of that magnitude have happened in the past multiple times, it still seems implausible that Ethereum could fly that high, doesn’t it? Maybe this time will be different and the peak will be much lower (or much higher). Nobody knows. I will personally be surprised if we don’t top $5000. That’s my low end target. For the high end, the biggest run-ups I’ve seen were order 1000x, so if it goes vastly more than $100,000 that would also surprise me. We shall see. Let’s talk recommendations:</p><p>First, <strong>I do NOT recommend buying large amounts of cryptocurrencies right now</strong>, near all-time highs. The right time to buy cryptocurrencies is when they are 5x to 20x down from a previous peak. Even when it is time to buy, only buy cryptocurrency with money you can afford to completely lose. No matter how painfully it drops, it can always go lower!</p><p>Second, <strong>NEVER sell all your cryptocurrency</strong>. No matter how ludicrously high it goes, it can always go higher.</p><p>Third, <strong>sell at regular price intervals</strong> as the price increases. I am selling some now, even though I expect and hope for much higher prices. To decide how much to sell, I try to imagine how much I will regret selling if the price goes up, and how much I will regret not selling if the price goes down. Everyone will have different pain thresholds in both directions. My current plan is to sell a third of my gains up to $5000/ETH, half of my gains up to $10,000/ETH, and a steadily increasing portion above that. If it gets as high as $100,000/ETH, I plan to have sold most of it by then.</p><p>Please understand: this is MY plan. It may not be appropriate for you. Talk to your financial advisor (and they will have a hearty laugh about it). Also, I should disclose that I am expecting that life circumstances will again force me to follow a different plan than what I described above, so I probably won’t even be able to take my own advice.</p><p>Still, I’m putting this out there now so I can hopefully say “I told you so” (yet again!) later.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ab443362bf76" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Time for me to Buy Cryptocurrencies (August 2018)]]></title>
            <link>https://medium.com/@jr.willett/time-for-me-to-buy-cryptocurrencies-august-2018-1d31cc1ecb8d?source=rss-1961ee6f2887------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d31cc1ecb8d</guid>
            <category><![CDATA[hype-cycle]]></category>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[market-timing]]></category>
            <category><![CDATA[bitcoin]]></category>
            <category><![CDATA[cryptocurrency]]></category>
            <dc:creator><![CDATA[J.R. Willett]]></dc:creator>
            <pubDate>Mon, 13 Aug 2018 23:34:01 GMT</pubDate>
            <atom:updated>2018-08-13T23:34:01.795Z</atom:updated>
            <content:encoded><![CDATA[<p>I have been involved with cryptocurrencies since 2010. Multiple times I have seen them spike up in a speculative bubble, only to collapse down to prices 90% off from their peak. The most recent price spike (in December/January) was driven largely by hype around Ethereum. I sold cryptocurrency during that hype, and I set cash aside, hoping to buy back during the subsequent crash. I decided I would start buying at 80% off the peak price, and be fully invested again if we hit 90% off.</p><p>Well, my friends, today Ethereum is more than 80% off its peak. I expect it to keep falling, but just in case, I put some cash back into cryptocurrency. I will buy more at 85% off, then again at 90% off. I plan to diversify into other coins as well during this buying opportunity.</p><p>I’m not trying to call the bottom. It could drop to 99% off peak for all I know. But I do know that I’ve seen this pattern several times before, and if the pattern holds, after all the weeping and bloodletting, the best projects eventually enter a new hype cycle and rise to new record heights that make the old highs look pitifully small. I’m not your investment advisor, and past performance can’t predict the future, etc. But I’m buying. If you have money you can afford to completely lose, now is definitely a better time to get into cryptocurrency than during the hype a few months ago, and even better (lower) prices may be just around the corner.</p><p>Please never put any money into cryptocurrency unless you can afford to completely lose it. Even if you don’t lose it all, the volatility requires a cast-iron stomach. Keep in mind that if you buy at 80% off peak, and the bottom is 95% off peak, you’ll have lost three quarters of your investment. I’m not advising YOU to buy. I’m just telling you what I am doing so I can say “I told you so” later. :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d31cc1ecb8d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Don’t Just Auto Invest! Proving Systematic Bias at LendingClub]]></title>
            <link>https://medium.com/@jr.willett/dont-just-auto-invest-proving-systematic-bias-at-lendingclub-c9a9122ec8c1?source=rss-1961ee6f2887------2</link>
            <guid isPermaLink="false">https://medium.com/p/c9a9122ec8c1</guid>
            <category><![CDATA[bias]]></category>
            <category><![CDATA[hedge-funds]]></category>
            <category><![CDATA[lending-club]]></category>
            <category><![CDATA[p2p-lending]]></category>
            <category><![CDATA[exploitation]]></category>
            <dc:creator><![CDATA[J.R. Willett]]></dc:creator>
            <pubDate>Thu, 19 Jul 2018 18:45:37 GMT</pubDate>
            <atom:updated>2018-07-30T16:54:16.389Z</atom:updated>
            <content:encoded><![CDATA[<p>I believe I can prove that <a href="http://lendingclub.com/">LendingClub.com</a>, the largest peer-to-peer loan platform, has for years been rating loans with systematic bias. Some borrowers are rated lower risk than they should be, and others are rated higher risk than they should be.</p><h3>Cause for Suspicion</h3><p>My investigation began with a simple question: Why do hedge funds get consistently high returns from peer to peer lending, while retail investors often report vastly lower returns using the same platform? For instance, the <a href="http://www.pmifunds.com/funds/">Prime Meridian Income Fund</a> keeps emailing me this graph of their fund’s absurdly good returns from peer to peer lending:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/889/1*tZNv5NnB60SUoVmzylh9Yw.jpeg" /></figure><p>How do they do this? According to them, they use “proprietary credit algorithms to overweight the best risk/reward loans”. In other words, at least some of the lending platforms do a poor job of rating their loans, and this fund developed an algorithm that rates loans correctly in order to exploit these inefficiencies.</p><p>So if the hedge funds are able to disproportionately stack their portfolio with loans that perform better than average, who gets stuck with the rest of the loans performing worse than average? That would be you and I, my friend: those of us who simply turn on the automatic investing, and let LendingClub diversify us across all the loans. Investing in all the loans will always lose compared to an algorithm that beats LendingClub’s own risk assessment methods, especially if rich hedge funds are scooping up all the best loans shortly after they appear.</p><p>It is easy to find <a href="https://www.lendacademy.com/my-quarterly-marketplace-lending-results-q1-2018/">people complaining about poor returns at LendingClub</a>, although it is difficult to prove anything from these anecdotes. So what can we prove? A lot, as it turns out, using <a href="https://www.lendingclub.com/info/download-data.action">LendingClub’s own publicly available data</a> (note: log in to your LendingClub investor’s account before downloading this data, otherwise some columns will be missing⁵).</p><h3>Loan Grading Basics</h3><p>LendingClub rates loans from A1 (the least risky, having the lowest interest rate) down to G5 (the most risky, having the highest interest rate). In general, loans rated A1 can be expected to default¹ slightly less than those rated A2, which default slightly less than those rated A3, and so on. This rating system basically works:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/977/1*Nnd3qMF6rDSTWSY3QP4Khg.png" /></figure><p>Things appear to break down when we get to the highest risk loans, but we have not yet accounted for some loans being paid off earlier than others. We will discuss early payoffs more later, but for now here is the same chart adjusted for the risk of early payoffs:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/977/1*sNfo8iKFIFiM9oeSHtYMTA.png" /></figure><p>Clearly the rating system basically works. However, we will find that many loans are several grades off from where they should be. Something rated B1 might actually be very low risk, and should have been rated A3. Something rated B3 might actually be much higher risk, and should have been rated C1.</p><h3>Mortgage or Rent?</h3><p>One of the biggest factors in most people’s credit score is whether they have a mortgage. Let’s start there. I’m sure LendingClub already takes this into account when they grade loans . . . right?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/977/1*DMS18OvJ9T-jQmq_Z_-c6g.png" /></figure><p>WHAT?! Renters default more than people with a mortgage in every single grade! Let’s take a closer look at the data. In the graph above, the effect appears to be smaller for the grade-A loans, but that is deceptive. A1 borrowers with a mortgage have a 1.45% chance of default, whereas A1 renters have a 2.19% chance of default. That means A1 renters are 51% more likely to default! The graph below illustrates this across all grades:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/977/1*FchRXlhUsqxLwKxKwF9Xtw.png" /></figure><p>Across all loan grades, being a renter increases the risk of default by an average of 26% compared to people having a mortgage. Before we get out the pitchforks, let’s see if LendingClub has fixed this. After all, we’re looking at more than 10 years of data. Perhaps this problem happened in older loans but is no longer an issue.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/856/1*BbZrbrY0rELfCfK9hkNwTg.png" /><figcaption>(Renters are on average 26% more likely to default, but this varies over time.)</figcaption></figure><p>Very few of the newest loans have defaulted, so that makes the signal noisy on the right side of the graph since there aren’t many defaults to analyze, but this doesn’t look good. It appears that LendingClub slightly over-penalized renters in 2008, but ever since 2010 it appears that renters have been under-penalized, and it might be getting worse. Hedge funds can massively increase their returns simply by manually selecting loans and avoiding renters. That’s incredibly low-hanging fruit.</p><h3>Unemployed People</h3><p>Having found that massive smoking gun, I went fishing for other inefficiencies. The next thing I found was that people who left their job title blank, or chose “N/A” for their length of employment tended to default much more than other loans. It is not a huge leap to assume that these people may be unemployed. Here are some more graphs, adjusted to remove rent bias²:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/1*YGG_4Qc6YofVaoU4_VS4ng.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/977/1*ZszRKgTOuQu9JPTJS_IeRg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/856/1*h7yAwoyub-Bv_2EQ3dhnsQ.png" /><figcaption>(Unemployed people are on average 34% more likely to default, but this varies over time.)</figcaption></figure><p>Note the Y-axis on that last graph above! It’s 4x wider than the graph for renters.</p><p>Surprise! Unemployed people are far, far riskier than other loans despite LendingClub lumping them in the same risk grade. It’s easy picking for the hedge funds to exclude these loans (and foist the unemployed on the rest of us). Again, the right side of the graph is noisy due to the loans being new and having very few defaults to analyze, but it appears this grading error may be getting worse over the past few months. Unemployed people who do pay off their loans also tend to pay them off fifteen days earlier on average, which makes them even riskier, as we will see later in the analysis.</p><h3>Credit History not Useful</h3><p>Next, I started looking at credit history. Unfortunately, the data provided by LendingClub does not show borrower credit history at the time they applied for the loan, but rather their <em>current</em> credit history⁵. Since their credit might be very different now from when they applied, the credit data provided by LendingClub can’t be used to show rating bias.</p><h3>Verification Status</h3><p>LendingClub provides a flag in their data called verification_status which they define as “Indicates if income was verified by LC, not verified, or if the income source was verified”. The three possible values are “Verified”, “Not Verified” and “Source Verified”. I’m not sure how they verify an applicant’s income, but one would generally assume that “Not Verified” would be an indicator of a higher risk loan. However, the data shows the opposite effect for some reason. These graphs show verified vs not verified after adjusting for rent bias and unemployment bias²:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/1*EbF5cuIOhuMtgtUovYj_tQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/977/1*IvfEjhW06BsJ4vYca8UKKg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/856/1*WxL_RhySaxGkfsLnHjEw5A.png" /><figcaption>(Verified people are on average 21% more likely to default than unverified, but this varies over time.)</figcaption></figure><p>Why unverified incomes would indicate a safer loan is an interesting question. Perhaps LendingClub automatically penalizes the rating of these loans, but does so too aggressively. Whatever the cause, it leaves more big piles of money for hedge funds to collect. They can take the people without income verification which have been mis-rated while the retail investor gets the less desirable loans who have verified income. Again, the data seems to show this problem getting worse in recent months.</p><p>With verification status, there is one big caveat. Loans categorized as unverified tend to be paid off earlier. Verified loans which are fully paid off after 605 days on average. Unverified loans are fully paid off after 537 days on average. This actually significantly increases the risk of unverified loans (see next section for an explanation). If we take into account early payoffs, the advantage of unverified loans over verified loans falls from 21% on average to 8%:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/977/1*ih8YbzpGZ0IgS-Nil-UAeQ.png" /><figcaption>Unverified loans still beat verified loans most of the time after considering early payoffs</figcaption></figure><p>Let’s take a moment to understand why early payoffs affect risk in this way:</p><h3>Understanding Early Payoffs</h3><p>Imagine two types of loans: loan type A is typically paid off in 1 year while loan type B is typically paid off early after 1 month. If they both have a 12% chance of default, then a portfolio of type A loans will have 1% of loans default every month, while a portfolio of type B loans will have 12% of loans default every month. They both have a 12% chance of default, but one type is 12 times riskier than the other! Many of the graphs here don’t account for early payoffs, but when early payoffs are a significant factor, I will provide additional information about its impact. (Early payoffs don’t appear to be different between renters and mortgage holders).</p><h3>Loan Purpose</h3><p>Next, let’s break down the data by loan purpose. Here’s a chart showing the loan purpose categories LendingClub uses and how many loans are in each:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/859/1*qciJl10KSv1CgPSWqgbKNg.png" /></figure><p>Next, let’s see how those categories perform versus the average loan (risks have been adjusted to account for rent bias, unemployment bias, and verification status bias²):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/859/1*S_c5s0kKcvz8YtPi-_u-1Q.png" /></figure><p>Here again early payoffs make a big difference. Here’s what the same chart looks like once we account for early payoffs:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/859/1*BvQf0y5vlgjJ7Z9Omx_ueg.png" /></figure><p>Clearly “educational” and “renewable energy” loans perform poorly, but there are almost none of them in the system so that isn’t very interesting. Weddings are typically not paid off early, which switches them from looking bad to looking good, but there aren’t many of those either. Instead, let’s look closer at car loans (lower risk) versus small business loans (high risk):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/950/1*dsMe9ni1i7_X_JazJN0HhA.png" /></figure><p>Once adjusted for risk of early payoff, A1 car loans have a 1.80% chance of default while A1 small business loans have a 4.87% chance of default: a 170% increase in risk! The advantage appears to go away for the riskier loans, but there are very few car loans in the riskier grades, so the data is noisy. Here’s the risk increase across all loan grades:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/860/1*J5t4X2FOtMZiBgO6c8INlg.png" /></figure><p>And over time:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/853/1*vWyXB9NzR-HRnIMJB753Mg.png" /></figure><p>Yup: the difference persists over time, and seems to get worse in recent months. Admittedly, car loans and small business loans won’t show up very often, but when they do, the hedge funds will scoop up the car loans and leave the small business loans for the retail investors (suckers) to pick up. I didn’t bother to make the graphs, but I expect that loans with listed purpose “other” are also good loans to grab when they appear.</p><h3>There Are No Bad Zip Codes</h3><p>Lending club provides the first three digits of borrower zip codes, and early passes at the data seemed to show biases by zip code, but in the end I concluded that LendingClub (or perhaps some data they rely on like a FICO score) is now accounting for zip code as best as can be done, starting in 2017. Prior to 2017 data, past performance of a zip code had some predictive value, but that value pretty much disappears in the latest loans. A word of caution to anybody else analyzing this data: since loans can last up to 5 years, an economic event in a particular zip code, such as a factory closing, can affect multiple vintages of loans. If the event happened in 2016, loans originating from as far back as 2011 could be affected. This can make it appear as if zip code past performance has more predictive value than it actually does. To test predictive power, you have to compare how zip codes from vintages more than five years earlier performed. Therefore I used zip code performance for 2010 loans to try to predict 2016 zip code performance, and 2011 loans to predict 2017 performance³ . Here are a couple graphs:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/853/1*CO-qSab4nPCfHDWdowYVmg.png" /><figcaption>Zip code past performance used to have some predictive power</figcaption></figure><p>A zip code which did poorly for loans originating in 2010 meant that same zip code would be slightly more likely to do poorly for loans originating in 2016.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/853/1*afPijFOau78c7reUa3bmCA.png" /><figcaption>No more predictive power</figcaption></figure><p>2017 zip code performance appears completely independent of how those same zip codes performed in 2011. After seeing so many things obviously done poorly, it is surprising to see zip codes being done so well. As far as I can tell, you can’t make a list of “bad zip codes” to avoid⁴.</p><h3>Smaller Loans (Under $5k) Outperform</h3><p>Next, let’s look at how default rates are affected by the size of the loan. Default rates have been adjusted to account for rent bias, unemployment bias, and verification status bias².</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*RdBjmN1rtoX93hebgjUlyQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*i5Hn7UCSBtpkWcuU-geOsw.png" /></figure><p>Loans under $5k are significantly less likely to default than other loans, perhaps because they are easier to repay? Let’s see how this holds up across loan grades:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/1*FLCxVaWWaXM6gLNm7mZAuw.png" /></figure><p>Loans under from $15k to &lt;$20k are 36% more likely to default than loans from $0 to &lt;$5k. The difference is fairly consistent across loan grades:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/1*Nh-RQWRBma3kQFf-M23Wrw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/856/1*PuQABr9NVMF7pAFjjHYfKw.png" /><figcaption>Loans $15k to &lt;$20k persistently default more than loans &lt; $5k over time, although the bias has been slowly getting better</figcaption></figure><p>By now you know what I am going to say next. The hedge funds are probably grabbing all these small loans while retail / auto investors get stuck with the larger loans which underperform.</p><p>Note the graphs above don’t account for early payoffs, which end up significantly changing some of the risks:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*_sxpd69AjK621IQtW1ImZQ.png" /></figure><p>When you account for the tendency for large loans to be paid off early, the biggest loans become the riskiest, while 25k to 30k loans become more attractive due to their tendency to be paid off later.</p><h3>Why Systematically Rate Loans Incorrectly?</h3><p>So why would LendingClub rate loans so poorly? I can speculate a couple reasons, which may or not be true:</p><ul><li><strong>They would rather have more hedge funds than more retail investors. </strong>The hedge funds contribute millions of dollars each while you and I contribute perhaps a few thousand dollars each, but they have to answer our questions and give us customer service too, even though we are a much smaller contributor to their profits.</li><li><strong>They can offer better rates to higher risk borrowers. </strong>The hedge funds supply capital to the credit-worthy borrowers, ensuring competitive rates for them, while many less credit worthy borrowers get BETTER than competitive rates due to the rating inefficiencies and the retail investing suckers who fund them via auto investing.</li></ul><p>Basically, it appears that everybody wins except the naive, trusting retail investor, who loses big time.</p><h3>Conclusions</h3><p>One really obvious conclusion is that if you have a LendingClub account with auto-investing turned on, you turn it off and invest manually or at least ADD FILTERS as I have done with my account:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*roeEaUTDZFSrJiQg1YYP3A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DPPC9GsKPC8kaIwrjW-Fxg.png" /></figure><p>You could also try moving to a different lending platform (which might have the same problem), or you could try to play hedge fund and choose better loans for yourself now that you are armed with the knowledge of the huge biases; the charts above indicate that a few simple rules could massively elevate your returns. The risk of playing hedge fund of course is that once a loan becomes available, anybody can fund it, and you will be racing against the hedge funds to grab the best loans. I expect the best loans are probably funded within milliseconds of becoming available. This is why I also plan to explore <a href="http://nsrinvest.com/">nsrinvest.com</a> to see if they might be able to increase my returns.</p><p>Another conclusion is that LendingClub should fix this problem. It’s hard for me to imagine that they don’t already know about it, since anybody half-decent with spreadsheets and/or databases could sniff out these biases (and many people have). However, I expect they have a lot of incentives to keep the biases in place as described earlier, so I’m only marginally hopeful they will take significant action to correct this.</p><p>Please note this is not an exhaustive list of biases, but simply the ones that were easiest for me to find. There are undoubtedly many more biases that hedge funds are exploiting.</p><h3>LendingClub Responds?</h3><p>I sent a draft of this article to LendingClub on July 3rd, more than two weeks before publishing it, and invited them to review my data and provide me with an official response. They have yet to reply. If they say anything to me, good or bad, I will update this section with any quotes from them.</p><h3>Methodology</h3><p>In order to generate these graphs, I imported all of LendingClub’s historical loan data into a Postgres database, and ran various hairy SQL statements to massage the data and generate statistical summary tables which I could then import into a spreadsheet where I massaged the data further and created graphs. My spreadsheet is publicly viewable <a href="https://docs.google.com/spreadsheets/d/1sJr0bcRRdFIIeC8ALAP9kWKE6J7JOC0Wc6s2BMZsm2Q/edit?usp=sharing">here</a>, although be advised it is NOT well-labeled or laid out for public consumption.</p><p>I’m just <a href="https://www.linkedin.com/in/jrwillett/">some guy</a> playing with a huge data set. I found and corrected dozens of errors in my analysis as I worked on it, and I was chasing errors up to the moments before publishing this article. Undoubtedly others will find more problems. Please let me know what problems you spot and I will try to correct them or at least acknowledge them in a footnote. I would <em>love </em>it if more people take a stab at independently verifying these findings.</p><h3>Other Analyses for Further Reading</h3><p>I am by no means the first person to find these patterns in LendingClub data. If you are looking for more interesting charts and analysis, here are a few:</p><p>This analysis from 2015 has some cool graphs, including showing the scandalous misrating of renters vs mortgage holders and of small business loans: <a href="https://rstudio-pubs-static.s3.amazonaws.com/115829_32417d32dbce41eab3eeaf608a0eef9d.html">https://rstudio-pubs-static.s3.amazonaws.com/115829_32417d32dbce41eab3eeaf608a0eef9d.html</a></p><p>This analysis using 2017 data also zooms in on the increased risk of renters vs mortgage holders: <a href="https://medium.com/@sonicmsba/how-to-minimize-risk-for-your-lending-club-investment-a4c8de0d129a">https://medium.com/@sonicmsba/how-to-minimize-risk-for-your-lending-club-investment-a4c8de0d129a</a></p><p>This analysis from all the way back in 2013 finds many of the same discrepancies I found: <a href="http://michaeltoth.me/analyzing-historical-default-rates-of-lending-club-notes.html">http://michaeltoth.me/analyzing-historical-default-rates-of-lending-club-notes.html</a></p><p>These brief analyses also point out the rent/mortgage discrepancy:</p><ul><li><a href="http://dfile.github.io/">http://dfile.github.io/</a></li><li><a href="https://www.orchardplatform.com/blog/2014522credit-variables-explained-home-ownership/">https://www.orchardplatform.com/blog/2014522credit-variables-explained-home-ownership/</a> (2012 data)</li><li><a href="https://nycdatascience.com/blog/r/p2p-loan-data-analysis-using-lending-club-data/">https://nycdatascience.com/blog/r/p2p-loan-data-analysis-using-lending-club-data/</a> (she didn’t think the difference was big enough to worry about!)</li></ul><h3>Footnotes</h3><ol><li>For the purpose of this analysis, “default” includes loans which are in default, charged off, or late more than 16 days. Including late loans allows us to see statistics on more recent loans, which haven’t had time to fully default even if the borrower has never made a payment.</li><li>Each new variable analyzed has been adjusted to remove the bias of previously analyzed variables, in order to make the variables more truly independent. For instance, for loan grade A4, there were 4619 unemployed people, of which 274 defaulted. However, 3.12 of those defaults can be attributed to unemployed people being more likely to be renters, so I calculated the risk of default for A4 unemployed people to be 270.88/4619 instead of 274/4619. Similar adjustments were made to employed people to account for their being slightly more likely to have a mortgage or own their home. I did not attempt to compensate for biases with more variables like loan purpose or loan size, as the calculations would have taken too long to complete.</li><li>When analyzing zip code prefixes, only those with more than 50 loans in a year were considered. Considering zip codes with fewer loans can make the data very noisy. This is why the graph for 2011 vs 2017 has so many more dots than 2010 vs 2016; the 2011 had far more zip code prefixes with more than 50 loans originated than 2010 did.</li><li>I actually have some graphs which appear to show that some zip codes do much better than others across all data, but this turned out to be hindsight bias: when you have a thousand candidates to choose from, it’s easy to find some who did better than others just by chance.</li><li>After I published this story, LendAcademy forum user Fred93 made an argument that the credit history is actually mostly from the time of application, despite some fields being newer. Unfortunately, LendingClub does not provide enough information to determine this from their description of the data, but Fred93 has been observing the data fo 10 years and believes the credit data is mostly not updated. Fred93 also alerted me to the fact that you must be logged in when downloading historical data or some columns will be missing. You can read the comments <a href="https://forum.lendacademy.com/index.php?topic=4892.0">here</a>.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c9a9122ec8c1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[UpToken Not Affected by BatchOverflow Bug]]></title>
            <link>https://medium.com/coinme/uptoken-not-affected-by-batchoverflow-bug-2fa32b05f797?source=rss-1961ee6f2887------2</link>
            <guid isPermaLink="false">https://medium.com/p/2fa32b05f797</guid>
            <category><![CDATA[batchoverflow]]></category>
            <category><![CDATA[ethereum]]></category>
            <category><![CDATA[bugs]]></category>
            <category><![CDATA[crypto]]></category>
            <category><![CDATA[uptoken]]></category>
            <dc:creator><![CDATA[J.R. Willett]]></dc:creator>
            <pubDate>Fri, 27 Apr 2018 17:37:28 GMT</pubDate>
            <atom:updated>2018-04-27T17:47:21.433Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rD3mvx2jkGwtKtRyFj0d6A.png" /></figure><p>Several Ethereum-based tokens were effectively destroyed this week by a smart contract bug called <a href="https://medium.com/@peckshield/alert-new-batchoverflow-bug-in-multiple-erc20-smart-contracts-cve-2018-10299-511067db6536">BatchOverflow</a>, but UpToken was not affected. Our decision to keep UpToken as simple as possible is paying off!</p><p>When we created the UpToken smart contract, we faced some criticism for being so conservative with our token design. We literally took the ‘minimum viable’ ERC-20 example code from Ethereum, and, without any alterations, <a href="https://etherscan.io/address/0x6Ba460AB75Cd2c56343b3517ffeBA60748654D26#code">used it to create UpToken</a>. We didn’t even include the usual code for handling a token sale — we filled all our orders manually.</p><p>In doing this, we went charging in the opposite direction of most of the market: everybody else seems to be trying to out-do each other with ever more elaborate smart contracts. While these contracts undoubtedly have useful and interesting functionality, on balance, they are short-sighted, and reflect a poor understanding of the risks inherent in having valuable assets managed by unchangeable code.</p><p>If you are designing a smart contract, keep this in mind: <strong>every single line of code is a potential vulnerability. </strong>By creating a smart contract to handle millions of dollars of transactions, you are creating a cash prize of millions of dollars to anyone who can break your contract. The best minds in the world will spend hundreds of hours poring over every line of your contract, trying to find some way to break in.</p><p>Think of it this way — if you had millions of dollars of gold in a vault, how many doors would you want into your vault? Every single line of code is an additional door that someone might figure out a way to force open. Yes, smart contracts can do sexy, amazing things, but you need to ask yourself: is all that sexiness really worth a hundred additional doors into my vault?</p><p>On top of that, if a year from now someone notices a crack in one of those doors, you can’t just replace it with a better door — in most cases you are stuck with the smart contract you created FOREVER.</p><p>Writing a complex smart contract to hold millions of dollars requires a lot of time, money, and humility. If you really must do this, you need to have your code carefully reviewed by other security experts, made available to the public for months before it goes into action, and ideally you should have big cash prizes available to anyone who can break it before it is deployed.</p><p>Make sure you understand the risks of using a smart contract. If you must use one, use the simplest, oldest, and most widely used code you can find, and resist the temptation to tinker!</p><p><em>This post was written by UpToken Senior Blockchain Engineer and </em><a href="https://www.forbes.com/sites/laurashin/2017/09/21/heres-the-man-who-created-icos-and-this-is-the-new-token-hes-backing/#3bc0de921183"><em>Founder of the ICO</em></a><em>, J.R. Willett.</em></p><p><em>For more information on UpToken, please follow us on </em><a href="https://twitter.com/UpToken"><em>Twitter</em></a><em> or join our </em><a href="https://t.me/UpToken"><em>Telegram Channel</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2fa32b05f797" width="1" height="1" alt=""><hr><p><a href="https://medium.com/coinme/uptoken-not-affected-by-batchoverflow-bug-2fa32b05f797">UpToken Not Affected by BatchOverflow Bug</a> was originally published in <a href="https://medium.com/coinme">Coinme</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>