<?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 Jairaj Sathyanarayana on Medium]]></title>
        <description><![CDATA[Stories by Jairaj Sathyanarayana on Medium]]></description>
        <link>https://medium.com/@jairajs?source=rss-4979423a427a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*opfwXD6kgNwvRcUx.jpg</url>
            <title>Stories by Jairaj Sathyanarayana on Medium</title>
            <link>https://medium.com/@jairajs?source=rss-4979423a427a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 04:46:26 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jairajs/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[What 36 months of Indian mutual fund holdings reveal about stocks, funds and embeddings]]></title>
            <link>https://medium.com/@jairajs/what-36-months-of-indian-mutual-fund-holdings-reveal-about-stocks-funds-and-embeddings-48c265df0de1?source=rss-4979423a427a------2</link>
            <guid isPermaLink="false">https://medium.com/p/48c265df0de1</guid>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[indian-equities]]></category>
            <category><![CDATA[word2vec]]></category>
            <category><![CDATA[mutual-funds-india]]></category>
            <dc:creator><![CDATA[Jairaj Sathyanarayana]]></dc:creator>
            <pubDate>Mon, 14 Sep 2026 16:16:01 GMT</pubDate>
            <atom:updated>2026-09-14T16:16:01.415Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Inspired by the ‘Asset and Investor Embeddings’ paperon 480,235 Indian fund-stock-month observations. Only active/ non-index funds were considered.</em></p><p>It started with the idea from <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4507511"><em>Asset and Investor Embeddings</em></a> by Xavier Gabaix, Ralph Koijen, Robert Richmond and Motohiro Yogo: instead of describing a stock only by its sector, financial ratios or text, learn about it from who owns it. If the same investors repeatedly hold two stocks together, that relationship should show up in a learned representation.</p><p>The first attempt used one month of Indian mutual fund holdings and a basic factor model. This version uses 36 months, from August 2023 to July 2026, and includes principal-component models, Word2Vec, portfolio and ownership transformers, simple baselines and proper held-out tests.</p><p>Here is the <a href="https://jairajs.github.io/india-mf-holdings-lab/">interactive explorer</a>. This is an explainer of sorts. <strong>Please note that this is strictly for informational/ fun purposes only, not investment advice, and accuracy is not guaranteed. The analysis plan was developed using a Grok Bot in an interactive manner and the final implementation used GPT-6 Astra.</strong></p><h3>1. Overview: start with the market map</h3><p>The Overview tab answers the most basic questions for any selected month: how many funds and stocks are covered, how broad the portfolios are, which stocks are held most widely, and where fund weights are concentrated.</p><p>In July 2026, the panel contains 248 funds and 895 stocks. The median fund holds 58 supplied stocks, and the disclosed equity positions add up to ₹30.7 lakh crore. That last number is not industry assets under management. It is the value of the equity holdings present in this dataset.</p><p>The breadth chart follows the same 171 funds across all 36 months, avoiding an obvious error where adding different funds over time creates a false trend. For this balanced group, the median number of stocks rises from 52 in August 2023 to 64 in July 2026. Portfolios appear to have broadened within the supplied panel.</p><p>The crowding table is equally useful. In July 2026, ICICI Bank is held by 160 funds, or 64.5% of the funds in that month. HDFC Bank is held by 150, Bharti Airtel by 142 and State Bank of India by 139. Eternal, formerly Zomato, is already held by 130 funds.</p><p>The ‘gaining’ and ‘losing’ tables should be read cautiously. A position appearing or disappearing can reflect a corporate action or reporting change, not necessarily a discretionary trade.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rc_intUvMrWUBkmIhhpP9A.jpeg" /></figure><h3>2. Stocks: look past the official sector label</h3><p>The Stocks tab lets you search any of the 1,080 securities, choose a month, and examine four different views: ownership, co-holdings, learned similarities and adjusted prices.</p><p>Take Infosys in July 2026. It is held by 120 funds, or 48.4% of the covered funds. Among funds that own it, the average weight is 2.64%. The value held within the panel is approximately ₹51,239 crore.</p><p>The Ownership view shows how the number of holders and average portfolio weight changed over time. It also lists every covered holder. This helps separate two different stories: more funds owning a stock, and existing holders increasing their allocations.</p><p>The Co-holdings view asks which stocks appear with Infosys in the same portfolios. Raw joint-holder counts are dominated by widely held stocks such as ICICI Bank and HDFC Bank. The more interesting measure is lift, which adjusts for that general popularity. Infosys and Wipro are held together about 2.07 times as often as independence would imply. Tata Consultancy Services is at 1.95 times, HCL Technologies at 1.89 and Tech Mahindra at 1.79.</p><p>The Learned similarities view is different. Word2Vec places Tata Consultancy Services closest to Infosys, which is intuitive. But the next few include Reliance Industries, State Bank of India, Axis Bank, Maruti Suzuki and Bharti Airtel. The representation is picking up portfolio roles and ownership patterns that do not remain inside sector boundaries.</p><p>That does not mean Infosys and Maruti have similar businesses, or that one can substitute for the other in a valuation model. <strong>It means they occupy partially similar positions in the portfolios used to train that month’s representation.</strong></p><p>You can also combine two stocks as a seed and find names close to their shared ownership theme. The Adjusted prices view displays the supplied monthly adjusted-close series. Prices do not train the embeddings. They are used separately in the return-comovement test described later.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IhzOAday3FosHCJA_wbacQ.jpeg" /></figure><h3>3. Funds: compare what managers own, not what the label says</h3><p>The Funds tab reverses the lens. It contains 258 active schemes across the full history and has three views: allocation and holdings, overlap and peers, and composition changes.</p><p>Consider HDFC Flexi Cap Fund in July 2026. The supplied file contains 75 stocks covering 94.3% of its net asset value (NAV). Yet its effective stock count is only 29.4. In plain English, the portfolio has 75 names, but its concentration is equivalent to roughly 29 equally weighted positions. Counting names alone would miss this.</p><p>Its largest supplied positions are ICICI Bank at 9.21%, HDFC Bank at 6.11%, Axis Bank at 6.01% and State Bank of India at 4.14%. The stacked chart shows how industry exposure changes over time. The grey remainder is deliberately labelled ‘outside supplied holdings’ rather than cash because the missing portion could include cash, debt, foreign securities or other out-of-scope assets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4TdSaPGSJrfzCj5zAbQxgw.jpeg" /></figure><p>The Overlap and peers view uses three measures. Weight cosine compares the pattern of allocations. Shared-name overlap measures how much the stock lists intersect. Shared NAV weight adds the smaller portfolio weight for every common position. It also shows transformer-based fund similarities as a separate lens.</p><p>For example, HDFC Flexi Cap and ICICI Prudential Large Cap have 35 common stocks, 28.2% shared-name overlap, 46.8% shared NAV weight and a weight cosine of 0.771. ICICI Bank is about 9.2% in both, but Reliance is 1.56% in the former and 5.75% in the latter. Two funds can therefore share many core names while expressing meaningfully different convictions.</p><p>The Composition changes view tracks entries, exits, shared-name overlap and total absolute weight change versus the prior month. For HDFC Flexi Cap, July 2026 has six new positions, one disappearance and 90.8% shared-name overlap. This is not portfolio turnover. Price changes, flows and corporate actions can all change reported weights without a corresponding manager trade.</p><h3>4. Research results: make the models beat simple alternatives</h3><p>The Research results tab is the core of the replication. I did not want a visually attractive embedding map without an out-of-sample test.</p><p>The main test hides the second-largest holding of a fund and asks each method to rank the missing stock using the rest of the portfolio. The complete test fund is excluded from training. Results are aggregated over six monthly test sets from February to July 2026.</p><p>Direct co-holding achieves 71.7% top-ten accuracy. Word2Vec reaches 70.3%. The portfolio transformer reaches 60.9%. A uniform random ranking would be expected to hit roughly 5.3% of the time.</p><p>The transformer did not win. A direct count of which stocks tend to occur with the visible portfolio was the strongest top-ten method.</p><p>There is some nuance. Word2Vec assigned better probabilities to the correct answer than co-holding, even though its top-ten hit rate was slightly lower. When an entire fund house was withheld in July 2026, Word2Vec reached 83.3% top-ten accuracy versus 78.4% for co-holding. The absolute percentages should not be compared directly with the six-month test because the test populations differ, but the result suggests some transfer beyond individual funds.</p><p>This is still not ‘stock-picking accuracy.’ The hidden target is the second-largest holding and is often a popular large-cap stock. The experiment measures whether portfolio structure can recover a missing position.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sYCFwKI3nCjmOKKcq_tnwA.jpeg" /></figure><p>I also reversed the task. For each test stock, I hid its second-largest covered fund owner and predicted that investor from the remaining ownership sequence. The ownership transformer achieves 53.7% top-ten accuracy, compared with 25.5% for simply ranking fund owners by popularity. Ownership structure contains signal in both directions.</p><p>The return-comovement result is less flattering, and more important to state clearly. A month’s holdings representation is used to explain the following month’s adjusted returns for held-out stocks. For each return month, the regression sees realized returns for 80% of the stocks and estimates returns for the other 20%. This tests whether the representation organizes stocks that experience similar shocks. It is not a trading forecast because the model uses contemporaneous returns from the reference stocks.</p><p>Industry classification explains 6.32% of held-out return variance. The best holdings representation, binary factors, explains 2.65%. Word2Vec explains 0.94%. My earlier one-month experiment suggested a stronger link between embedding similarity and return correlation. That conclusion did not survive the broader and cleaner test.</p><h3>5. Data and methods: know what the panel cannot tell you</h3><p>The Data and methods tab documents the source, definitions, model choices, validation and limitations.</p><p>The panel, collated by <strong>Aaditya Porwal</strong> from individual fund-house websites, contains 480,235 fund-stock-month rows covering 11 fund houses, 258 active schemes and 1,080 securities. Adjusted prices come from bhavcopy at the supplied adjusted-close dates.</p><p>There are no duplicate fund-stock-month keys or conflicting prices for the same stock-month. There are, however, 2,322 positive positions whose reported portfolio weight rounds to zero, 423 missing stock-month prices out of 29,442, and a complete UTI coverage gap in October 2025. Positive shares and market value are used to retain rounded-zero positions. Missing months and prices are never filled with zeros or forward-filled.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D2zEUGyqy97UQU3sglmewQ.jpeg" /></figure><p>For the learned representations, each monthly fund-stock network is filtered until every included fund has at least 20 stocks and every included stock has at least 20 holders. In July 2026, that leaves 227 funds and 240 stocks. The models use 32-dimensional representations. Factor and Word2Vec models are fitted monthly; transformers are fitted for the final six months.</p><p>The principal-component variants use binary ownership, within-fund ranks or log market values. Word2Vec treats a portfolio like a sequence of stocks ordered by weight and learns which names occur in similar contexts. The transformers mask positions and learn to reconstruct them. No pretrained language model, company fundamentals or generated economic narrative enters the results.</p><p>For the hidden-holding tests, funds are assigned deterministically to training, validation and test groups. Test funds never enter model fitting. The return test uses a separate fixed split of stocks and selects its regression settings using only the training stocks.</p><p>The dataset does NOT contain company market cap, book equity, accounting fundamentals, earnings-call text, complete fund NAV returns or verified disclosure timestamps. I therefore did not reproduce the paper’s relative-valuation, accounting, text or trading-strategy tests. There is no alpha claim here.</p><h3>What I would use this for</h3><p>The useful output is a new map of the market built from revealed portfolio choices. It can help generate peer sets that cross sector boundaries, identify crowded ownership, compare funds beyond their category labels and surface portfolio changes worth investigating.</p><p>It should be an input into research, not the conclusion. Holdings similarity can reflect mandate, size preference, liquidity, benchmark awareness or fund-house policy. It does not by itself establish business similarity, mispricing or expected return.</p><p>The broader lesson was simpler: Indian (active) mutual fund holdings contain a surprisingly coherent map of stocks and investors, but simple co-holding already extracts much of it. Deep models add information in some tests. They are not automatically the answer.</p><p>Explore the <a href="https://jairajs.github.io/india-mf-holdings-lab/">full site here</a>.</p><p><strong>Strictly for information and fun. Not investment advice. Accuracy is not guaranteed. The analysis, website, this blog post were all prepared using AI and should be independently checked.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=48c265df0de1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>