1,777 insider purchases, one calendar test, and a 5.3-point gap that fourteen years of publicity still hasn’t closed.

Same kind of trade, opposite results: the buys that broke the calendar beat the S&P 500, the routine ones trailed it. Source: SEC Form 4, WallStPal.

Every March, like clockwork, a certain kind of executive buys a little more of their own company’s stock. You could set your calendar by it. And according to a well-known study and the 1,777 purchases I just tested, you can ignore almost every one of those buys.

The buy worth your attention is the one that breaks the habit.

In Part 1 of this series I argued that insider buying matters and insider selling mostly doesn’t. This part goes back to buying, and to a problem hiding inside it.

A lot of insider purchases aren’t decisions at all. They’re habits.

The trick that fits on a napkin

In 2012, three researchers, Lauren Cohen, Christopher Malloy and Lukasz Pomorski, published a paper in the Journal of Finance called “Decoding Inside Information.” Their idea was simple enough to fit on a napkin.

Look at each insider’s own history. If someone has traded in the same calendar month for several years in a row, every March for example, call them a routine trader. Their trades follow a calendar, not a view about the company. Bonuses get paid, trading windows open, and the trade happens.

Everyone else with enough history is an opportunistic trader. They trade when they decide to, and the timing is the information.

The gap they found was large. Routine trades predicted essentially nothing. A portfolio built on opportunistic trades earned about 0.82% a month above the market, value-weighted, and about 1.8% a month equal-weighted. Opportunistic trades also came before company news, like earnings surprises, while routine trades did not.

There was one more finding that I think about a lot. Opportunistic trading was linked to a higher chance of SEC enforcement later. Routine trading was not.

Fourteen years later, does it still work?

Fourteen years is a long time for a published edge to survive. Once everyone knows a trick, it usually stops working. So I ran the same split on the insider purchases in the WallStPal database.

The rule I used: an insider who traded in each of the last three years is routine if they also traded in this same calendar month in each of those years, and opportunistic if they didn’t. Anyone without three years of history is left unclassified. The trading history comes from SEC Form 4 filings and the SEC’s quarterly insider data sets.

Then I measured each purchase the same way every time: the stock’s return over the next three months, minus the S&P 500's.

Two groups doing the same thing, buying their own company’s stock, and a 5.3-point gap between them. Source: SEC Form 4, WallStPal.

The split still works. Opportunistic purchases beat the market by a median of 1.4 points. Routine purchases trailed it by 3.9 points. That’s a gap of 5.3 points between two groups of people doing what looks like the same thing: buying their own company’s stock.

The routine number is the surprising one. Not only did routine buys fail to beat the market, they did worse than the purchases we couldn’t classify. Buying on a schedule is not a signal of anything.

One honest note on the selling side. The same split barely matters for sales: opportunistic sales were followed by −2.5% against the market, routine sales by −3.7%. Selling stays noisy however you slice it, and the next part of this series is about the few sales that do matter.

Spot one yourself with two questions

Three Marches in a row is a habit. The November buy is the one telling you something.

You don’t need software for this. Open an insider’s filing history on the SEC’s EDGAR site or on WallStPal and ask two questions.

1. Has this person traded every year? If they only bought once, ever, you can’t tell routine from opportunistic. That’s most insiders.
2. Do the trades bunch in the same month? If they bought in March of 2023, 2024 and 2025, and this purchase is in March again, it’s probably routine. If they bought in March three years running and now they’re buying in November, that’s the one to read.

Who’s breaking the calendar right now

The opportunistic purchases filed in the last few months are a mixed bag, and it’s worth being honest about why.

A large part of the list is not executives at all. It’s funds that own 10% or more of a company, or that have a partner on its board. Berkshire Hathaway bought about $400 million of Lennar stock over seven trading days in late September. Baker Bros. bought Kodiak Sciences and Madrigal Pharmaceuticals. OrbiMed and RA Capital bought into several young biotech companies where they have a partner on the board, which often looks more like financing than a view on the price. A fund has its own reasons and its own calendar, so a fund on this list deserves a different reading than a CEO.

The executive purchases are the ones closest to what the research measured:

• Uber (UBER). CEO Dara Khosrowshahi bought $10.0 million of stock on September 10, 2026.
• Intel (INTC). CEO Lip-Bu Tan bought $10.0 million on August 11, 2026 and another $9.7 million on November 2.
• MGM Resorts (MGM). Director Keith Meister bought $51.5 million on October 30, 2026.

None of these is old enough for the three-month test yet, so this is a list to watch, not a scorecard.

The quiet buyer who may know more than the CEO

The research found the most informative opportunistic traders were often not the CEO. They tended to be less senior insiders, at companies near where they lived, with weaker governance. A $10 million CEO purchase gets the headlines. A $200,000 purchase by a regional vice president who has never bought in that month before may say more.

The gap between opportunistic and routine buys is a median. Plenty of opportunistic purchases lost money: 46% of them trailed the market. One trade, however well timed it looks, can still go wrong. This is research, not investment advice.

The live list, with the track record updated every day, is at wallstpal.com/opportunistic-buys. There’s a longer explainer at wallstpal.com/guides/opportunistic-insider-trading.

Sources: WallStPal opportunistic insider buying, data as of November 9, 2026; Cohen, Malloy and Pomorski, “Decoding Inside Information,” Journal of Finance, 2012 (summary); SEC Form 4 filings.

Make Better Financial Decisions With Your Own Numbers

At Investor’s Handbook, we turn real investing and money questions into practical frameworks and free downloadable Excel models you can use to evaluate your own decisions.

Join our free newsletter for the Tuesday Digest, Thursday Weekly Framework, and a growing library of financial tools covering investing, retirement, debt, and building wealth.

📩 Get Free Financial Models & Join Investor’s Handbook

Learn the framework. Run the model. Make the decision.

🎥 Watch on YouTube · 𝕏 Follow on X

Insiders Buy Their Own Stock Every Spring. Watch the One Who Suddenly Buys in November. was originally published in Investor’s Handbook on Medium, where people are continuing the conversation by highlighting and responding to this story.

“CEO dumps $25 million in stock.” You have seen that headline. Maybe you sold because of it. Maybe you went looking for your own Big Short.

Here is what the people who study this for a living keep finding: that headline almost never means what you think. The sell column is mostly noise. The real signal sits in the column almost nobody reads, the one where an executive spends their own cash on their own company’s stock.

Tesla’s filings from the past year show it in plain sight.

Tesla: $87.7 million of selling, and the one trade that mattered

Start with what the headlines would show you. Over the past 12 months, Tesla insiders filed 10 sales worth $87.7 million and not a single open-market purchase (WallStPal, Tesla insiders). Four insiders, all selling.

Sounds ominous. Now read the fine print. Four of the ten sales ran on what is called a 10b5–1 trading plan.

Think of a 10b5–1 plan as an autopilot order. At a time when an insider holds no inside information, they give their broker written instructions: sell this many shares, on these dates or at these prices. From then on they cannot steer the trades, and since 2023 the rules make officers and directors wait at least 90 days before the first sale. So when a plan sale shows up in the news, the decision behind it was made months earlier, often for reasons like taxes or diversification.

The CFO’s sales are small, nearly identical in size and spaced about a quarter apart. That is a schedule too. Nobody in this table is telling you anything about Tesla’s future.

Now look just outside that 12-month window, at the trade most people have already forgotten. On September 12, 2025, Elon Musk bought about 2.57 million Tesla shares in 25 trades at $372 to $396 each. Roughly $1 billion of his own money (Gulf News). When the filing hit on September 15, the stock jumped as much as 9% before the open.

Ten sales on autopilot, or one CEO writing a billion-dollar check. Which one tells you what the person with the most information actually thinks?

Now the part most insider-trading articles skip. Musk paid about $389 a share. Tesla closed October 2025 at $456.56, about 17% above his price. By November 2026 it was at $341.38, about 12% below it (Digrin). The person who knows more about Tesla than almost anyone spent $1 billion, and fourteen months later he was underwater on paper.

A purchase tells you what an insider believes. It does not make them right. Insiders misjudge the economy, their competitors and their own timing like everyone else. Treat a buy as evidence, never as a guarantee.

Why the sell column lies to you

Think about how executives get paid. Most of their wealth arrives as stock in one company. To buy a house, pay a tax bill, fund a foundation or simply stop having every dollar in one basket, they have to sell. None of that says a word about next quarter.

Three things turn most insider sales into background noise:

• The plan was set months ago. Rule 10b5–1 lets insiders schedule sales in advance. Since the SEC tightened the rule in December 2022, directors and officers must wait at least 90 days, and up to 120, before a new plan can trade (SEC fact sheet). The sale you read about today was decided last season.
• It happens every year. Cohen, Malloy and Pomorski found that routine trades, made in the same month year after year, are more than half of all insider trades. A portfolio built on them earned essentially nothing extra (NBER).
• Rich people sell big. Since August 2025, NVIDIA’s Jensen Huang sold $702.7 million of stock and Arista’s Jayshree Ullal sold $1.51 billion, every share under a trading plan (WallStPal). Huge numbers, almost no information.

Peter Lynch said it best in One Up on Wall Street: insiders sell for many reasons, “but they buy them for only one.” They think the price is going up.

The evidence academics found 25 years ago

This is not a hunch. Three of the most cited studies on insider trading, published over more than a decade, keep landing in the same place: the money is in the purchases.

Why would buying be so different? Because a purchase hurts. The insider spends their own cash, piles more of their wealth into one stock and puts their judgment on public record. Nobody does that by accident. A sale has a dozen boring explanations. A purchase has really only one.

Two trades that tell the whole story

The buy: Jamie Dimon, February 2016

Early 2016 was ugly for banks. A global sell-off had JPMorgan down almost 20% for the year and about 25% below its July high. Investors were running.

Jamie Dimon walked the other way. He bought 500,000 shares at an average of $53.18, about $26.6 million, roughly what he had been paid for all of 2015 (CNN Money). Within ten months the stock had rallied nearly 55% (Yahoo Finance). By the last trading day of 2016 it closed at $86.29, 62% above his price (Digrin).

The sale: Greg Becker, February 2023

So is every sale noise? No, and this is the case that proves it. On February 27, 2023, Silicon Valley Bank’s CEO sold 12,451 shares for $3.6 million under a plan he had set up on January 26. It was his first sale in more than a year. Eleven days later the bank collapsed (Fortune).

Look at what made it different. He broke his own pattern. The plan was created only weeks before the bad news. Those are the fingerprints of a sale that means something, and they are exactly what the SEC’s 90-day waiting period, in force from 2023, was built to stop. A sale that repeats every quarter at the same size carries none of them.

So why isn’t everyone rich from this?

If the secret is that simple, just buy what insiders buy. Plenty of smart people tried. Here is what happened to them.

The funds that died. The Guggenheim (later Invesco) Insider Sentiment ETF, ticker NFO, bought 100 stocks where top executives were buying and analysts were raising estimates. Direxion launched a rival, KNOW, in December 2011 (ETF.com). NFO shut down in February 2020. KNOW followed in October 2020. CXO Advisory tracked nine guru and insider ETFs from launch through September 2026: together they grew 10.3% a year, while the S&P 500 grew 11.3% (CXO Advisory).

The filter that worked. The academic strategies that won did not buy every purchase. Cohen, Malloy and Pomorski’s 82 basis points a month appeared only after they threw out the routine traders. The edge lives in the filter.

What copying looked like in 2026. WallStPal measured all 7,071 open-market purchases filed in 2026, worth $32.9 billion. One month later, 57% were beating the S&P 500. Three months later, only 41% were, and the median trailed the index by 3.7 points. Directors’ purchases held up best. Buys by two or more insiders within 30 days did a little better than the rest (WallStPal).

So here is the full secret, including the part most articles leave out. Insider buying has carried real information for decades. The edge is narrow, it fades within months, and in a year when a few giant stocks carry the whole index, it can disappear. Treat a purchase as a flashlight that tells you where to look. It is not a buy button.

Read the filings like an insider

Every one of these trades is public. Insiders must file a Form 4 within two business days, and anyone can read it. Before you react to the next headline, run six checks:

1. Read the code. “P” is an open-market purchase and “S” is a sale. Grants, option exercises, tax withholding and gifts use other codes and mean little.
2. Find the plan footnote. A sale marked as made under a 10b5–1 plan was scheduled months earlier.
3. Check the pattern. Does this person trade in the same month every year? Routine trades carry almost no signal.
4. Size it against the stake. A $1 million buy from someone who owns $2 million of stock says far more than the same buy from someone who owns $2 billion.
5. Look for company. Two or more insiders buying within a few weeks is a stronger sign than one.
6. Read the setting. A buy after a steep fall, like Dimon’s in 2016, means an insider is betting against the crowd.

So where is the Big Short of 2026? It is not in the sell column, where ten Tesla sales told you less than one purchase did. The next time a headline screams that a CEO is dumping stock, open the filing, skip to the buy column, and ask one question: who is spending their own money, and why now?

And remember Tesla’s last lesson. Even a $1 billion purchase is a bet, and bets can lose.

This article is for information only and is not investment advice.

Sources

• Gulf News: Elon Musk buys $1 billion worth of Tesla stock
• WallStPal: Tesla insider trades
• SEC: Rule 10b5–1 amendments fact sheet
• Cohen, Malloy & Pomorski, Decoding Inside Information (NBER)
• Lakonishok & Lee, Are Insider Trades Informative? (RFS 2001)
• Jeng, Metrick & Zeckhauser, Estimating the Returns to Insider Trading (REStat 2003)
• WallStPal: CEOs who never buy
• CNN Money: Jamie Dimon buys $26.6 million in JPMorgan stock
• Yahoo Finance: Trade of 2016, Dimon buys at the bottom
• Digrin: JPMorgan and SVB monthly price history
• Digrin: Tesla monthly price history
• Fortune: SVB CEO Greg Becker’s $3.6 million stock sale
• ETF.com: 2 insider-buying ETFs, very different returns
• CXO Advisory: Do ETFs following gurus and insiders work?
• WallStPal: Insider buying vs the S&P 500 in 2026

Make Better Financial Decisions With Your Own Numbers

At Investor’s Handbook, we turn real investing and money questions into practical frameworks and free downloadable Excel models you can use to evaluate your own decisions.

Join our free newsletter for the Tuesday Digest, Thursday Weekly Framework, and a growing library of financial tools covering investing, retirement, debt, and building wealth.

📩 Get Free Financial Models & Join Investor’s Handbook

Learn the framework. Run the model. Make the decision.

🎥 Watch on YouTube · 𝕏 Follow on X

The Big Short 2026 was originally published in Investor’s Handbook on Medium, where people are continuing the conversation by highlighting and responding to this story.

Sponsored by Snowball Analytics.

Part 1 of 4 in a series on using AI with real portfolio data. Part 2: Can ChatGPT Analyze Your Portfolio? I Tested a Screenshot Against Real Data · Part 3: How Much Will My Portfolio Pay Me Next Year? · Part 4: The 6 AI Portfolio Review Prompts I Run Every Month

I asked ChatGPT how my portfolio did this year. It answered with percentages, sectors and a short summary, confidently, without having seen a single one of my positions. The model knows a lot about markets and nothing about my account.

A pasted screenshot doesn’t fix that. It shows one screen on one day, with no transactions, cost basis or dividends, and only one broker. So I gave ChatGPT and Claude read-only access to the portfolio data over MCP, the standard way an AI app reaches an outside data source, and checked their answers against the dashboard.

Here’s how to set it up, what the AI can see, and what it can’t touch.

Why access matters more than intelligence

The model is perfectly capable of explaining a portfolio; what it lacks is a structured source of truth to answer from. With MCP, the AI calls the tracker, the tracker does the math, and the AI explains the result in plain language.

Three ways to ask AI about your real portfolio (cheapest first)

Snowball Analytics is a portfolio tracker that connects to 1,000+ brokers across 70+ exchanges, so all the holdings sit in one place. There are three ways to put AI on top of that data.

For every test in this article I used Snowball’s Demo portfolio rather than my own account: nine positions plus cash, with real Snowball calculations on sample holdings.

1. Ask Snowball Analytics, inside the product

This is the simplest option and it works on every plan, on web and mobile, with usage limits that depend on the plan. You type a question inside Snowball and get an answer based on the holdings, with no setup and no keys.

My first question was: “What are my top 5 holdings by weight?”

Ask Snowball Analytics, the assistant inside the product.

The answer came back as a table: AMD 58.25%, NVDA 20.98%, META 5.47%, VOO 4.89% and AAPL 4.73%. It also flagged the obvious problem on its own: AMD alone is more than half the portfolio, and the two chip stocks together are about 79%.

Best for: quick questions while you’re already in the app.

2. ChatGPT, connected over MCP

If you already live in ChatGPT, you can connect it directly to your Snowball data. MCP access is part of Snowball’s Expert plan (plan details are on the pricing page). I tested it on ChatGPT Plus, adding Snowball as a custom MCP server, and you call it inside any chat by typing @snowball-analytics before your question.

ChatGPT (Plus) answering through the Snowball connector, called with @snowball-analytics.

Best for: people who already work in ChatGPT and don’t want to switch.

3. Claude, connected over MCP

The same connection works in Claude, on the web, in the desktop app and in Claude Code, the terminal version. The difference from the in-app assistant is context: Claude can combine the portfolio data with everything else you’re doing in the conversation, like drafting a review, structuring a monthly checklist or explaining a term.

Best for: longer analysis sessions and repeatable routines.

How to connect Snowball Analytics to ChatGPT and Claude, step by step

The setup takes a few minutes. The first part is the same for every app: get your key. Then you follow the steps for the app you use. Snowball’s own guide is in the help article.

Part 1: Get your key in Snowball (all apps)

Step 1 — Open your profile and go to the MCP tab.
It sits next to Account, Subscription settings and Display settings. You need the Expert plan: without it, the tab explains the feature but shows no button.

Step 2 — Click Generate.
Snowball shows two things: your key and the MCP server address. The key is shown once, so copy it somewhere safe before leaving the page. If you lose it, click Regenerate and the old key stops working right away.

The MCP tab in my Snowball profile. The key is blurred.

Treat the key like a password. Don’t paste it into a chat, and blur it in any screenshot.

Which address to use. Always use the server address shown on your MCP tab. There are two ways to pass the key with it:

• With a header: the server address plus a header X-API-Key with your key. Only some apps let you add headers.
• Inside the address: the same server address with /k/ and your key added at the end, and authentication set to none. This works in apps that only ask for a URL, which is the case for the Claude app and ChatGPT.

Part 2a: ChatGPT

I did this on ChatGPT Plus. In ChatGPT, external MCP servers live in the Plugins section of Settings (my ChatGPT is in Spanish, where it’s called Complementos).

1. Open Settings → Plugins and choose to add a custom MCP server.
2. Name: snowball-analytics. The icon and description are optional.
3. Connection: leave Server URL selected (not Tunnel) and paste the address from your MCP tab with /k/ and your key at the end.
4. Authentication: ChatGPT pre-selects OAuth. Since the key already travels in the address, Snowball doesn’t need a login, so set it to None.
5. Read the warning about custom MCP servers, tick I understand and want to continue, and save.
6. In a new chat, type @snowball-analytics followed by your question. The @ tells ChatGPT to use the connector for that message.

Adding Snowball as a custom MCP server in ChatGPT (Spanish interface). Key blurred.

Once saved, the app’s detail page confirms the setup: authorization None, version dev mode. ChatGPT treats custom MCP servers as apps in development, built by you.

The connected Snowball app in ChatGPT. Key blurred.

One thing I didn’t expect: once the connector was installed, ChatGPT reached for it on its own, even in a chat where I hadn’t typed the @. When I later wanted an answer without the portfolio data, I had to say so in the prompt (“don’t use any apps or connectors”). The warning ChatGPT shows when you add a server is also fair: connect only servers you trust. Snowball’s connection is read-only, so it can’t modify or delete anything in your portfolio.

Part 2b: Claude (web and desktop app)

1. In Claude, open Settings → Connectors. The Yours tab lists the connectors you already have.
2. Click + Add (top right) and choose to add a custom connector.

Settings → Connectors in Claude. Custom connectors are added from + Add.

3. In Add custom connector, type a name (snowball-analytics) and paste the address from your MCP tab with /k/ and your key at the end. There's no OAuth step and nothing else to fill in.

4. Click Continue.

Adding Snowball as a custom connector in Claude. Key blurred.

5. Open the new connector. Claude lists Snowball’s read-only tools, and by default each one is set to Needs approval, so Claude asks before every call. If you’d rather not click “allow” every time, you can switch them to always allowed. Since they’re all read-only, the risk is low, but the choice is yours.

The Snowball connector in Claude: read-only tools, each needing approval by default. Key blurred.

6. In a new chat, ask your question. Claude decides on its own which Snowball tool to call. With the default setting, it stops and asks first: Deny, Always allow or Allow once.

Claude asks before calling a Snowball tool. Its first move: list the portfolios.

Part 2c: Claude Code (terminal)

In Claude Code, the whole setup is one command. Use the address from your MCP tab with /k/ and your key at the end:

claude mcp add --transport http --scope user snowball-analytics <ADDRESS_FROM_YOUR_MCP_TAB>/k/YOUR_KEY

Then claude mcp get snowball-analytics confirms the connection:

Snowball Analytics connected to Claude Code. Key blurred.

Restart Claude Code after adding it, because the tools load when a session starts.

Part 3: Ask your first question

Ask something only your data can answer, like “What are my top 5 holdings by weight?” If the app asks you to sign in to Snowball, the key isn’t being sent, so check the address. Snowball never asks for your password inside an AI app. If you track more than one portfolio, ask for the list first and then name the one you mean.

First real test: does the answer match the dashboard?

A connection is only useful if the numbers are right. So I asked Claude the same question I asked inside Snowball:

“What are my top 5 holdings by weight?”

Claude listed the portfolios, called Snowball’s tools and answered:

Claude’s answer, using the snowball-analytics connector.

Together, those five are about 94% of the portfolio. Claude went a step further without being asked: AMD and NVIDIA together are almost 80% in two semiconductor stocks, which explains a beta of about 2.2.

Against ChatGPT: the answer was identical, with the same five holdings in the same order and the same weights to the second decimal.

Against Ask Snowball: same holdings and order. Four weights match exactly; AMD shows 58.25% in the app and 58.26% through MCP, a rounding difference of 0.01 points. The other difference is currency: the in-app assistant answered in euros (€376,769 for AMD) and Claude in dollars ($423,384), so if you want a specific currency, say so in the question. All three assistants gave the same answer because all three read the same source.

Then I ran the check that matters most, against the Snowball dashboard itself.

The holdings view in the Snowball dashboard, in euros.

The holdings view doesn’t show weights, and it shows euros, so I checked two things the AI can’t fake:

• Cost basis. AMD €335,994.11, NVDA €128,010.35, META €30,624.74, VOO €12,739.25, AAPL €8,622.87. Identical to the cent to what Ask Snowball returned.
• Relative size. If the weights are right, the gaps between positions must match the dashboard, whatever the currency. Within these five, AMD is 61.76% of the dashboard’s market value, and 61.76% using Claude’s weights. NVDA: 22.24% and 22.24%. META, VOO and AAPL: 5.80%, 5.18% and 5.02% on the dashboard, 5.80%, 5.18% and 5.01% from Claude.

The only gap is one hundredth of a point on Apple, which is rounding. The numbers line up because Claude isn’t doing the math: it calls Snowball, Snowball calculates, and Claude writes the answer in plain language.

Is it safe to give AI access to your portfolio?

This is the first question everyone asks, so here’s exactly what the connection can and can’t do.

It’s read-only. Nothing can be bought, sold, added or edited from a chat. The AI can read your portfolio data, but it can’t touch it.

The key is under your control. It’s generated on the MCP tab of your profile, shown once, and stored as a hash. If you want to cut access, you revoke it with Regenerate or Delete.

The AI doesn’t compute your numbers. Every figure is calculated by Snowball Analytics. The AI app only formats and explains what it receives.

What the AI app receives. The data sent is anonymized, and Snowball Analytics doesn’t use it to train models. The AI app itself (Claude, ChatGPT or any other) follows its own data settings, so check the privacy settings of the app you connect. No broker passwords are shared with the AI, and no orders can be placed.

What the AI can actually look up

Snowball’s MCP tools are all read-only. You never call them by name: you ask in plain language, and the AI picks the right one. Knowing what’s there tells you which questions are worth asking.

These are the ones I use most:

The rest cover your transaction history (GetTransactionList, GetTransactionSummary), the list of your portfolios (ListPortfolios), and research on any company or fund, not only the ones you own (GetCompanySummary, GetCompanyFundamentalsSummary, GetCompanyPriceReturns, GetFundSummary).

Upcoming payouts from GetPortfolioEvents also show in the dividend calendar.

FAQs

Can ChatGPT see my brokerage account?
Not by default, because a chatbot only sees what you give it. With an MCP connection to Snowball Analytics, it can read your portfolio data (read-only), but it never gets your broker login.

Can the AI place trades or change my portfolio?
No, because the Snowball MCP connection is read-only. Nothing can be bought, sold, added or edited from a chat.

Do I need the Expert plan?
For ChatGPT, Claude or other MCP apps, yes: MCP is part of the Expert plan. Ask Snowball Analytics, the in-app assistant, works on every plan.

Your next step

Generate a key on the MCP tab, add it to the AI app you already use, and ask the same question I started with: “What are my top 5 holdings by weight?” Then open the allocation view in the dashboard and compare. If the numbers match, you have a source of truth you can question in plain language.

Next in this series: Can ChatGPT Analyze Your Portfolio? I Tested a Screenshot Against Real Data. From a screenshot, two of three answers were off by 4x to 37x.

Try Snowball Analytics free for 14 days, no credit card. Registration comes with a 14-day free trial that includes paid features.

Try Snowball Analytics free for 14 days

This article is for information and education only and is not investment, tax or legal advice. Snowball Analytics SAS is not a broker-dealer, bank or registered investment adviser. AI answers describe the portfolio data used in this article (Snowball’s Demo portfolio); they are not recommendations. Projected income is an estimate and may differ from actual payments. Past performance does not predict future results.

Make Better Financial Decisions With Your Own Numbers

At Investor’s Handbook, we turn real investing and money questions into practical frameworks and free downloadable Excel models you can use to evaluate your own decisions.

Join our free newsletter for the Tuesday Digest, Thursday Weekly Framework, and a growing library of financial tools covering investing, retirement, debt, and building wealth.

📩 Get Free Financial Models & Join Investor’s Handbook

Learn the framework. Run the model. Make the decision.

🎥 Watch on YouTube · 𝕏 Follow on X

How to Connect ChatGPT and Claude to Your Portfolio (Read-Only, Step by Step) was originally published in Investor’s Handbook on Medium, where people are continuing the conversation by highlighting and responding to this story.

Every February, May, August and November, the same screenshots go around. Warren Buffett bought this. Michael Burry sold that. Bill Ackman opened a new position. The posts collect thousands of likes, and plenty of people add those stocks to a watchlist or straight to their account.

I understand the appeal. These are some of the most successful investors alive, their holdings are public, and following them feels like borrowing their homework.

I run WallStPal, a site that reads SEC filings every day, so I had the data to check whether borrowing that homework pays. I took 28 of the most followed funds, from Berkshire Hathaway and Bridgewater to Pershing Square and Carl Icahn, and copied their new buys the way an ordinary investor would have to. Then I compared every pick with putting the same money into the S&P 500 on the same day.

Two funds out of 28 came out ahead. For the other 26, a plain index fund would have done better.

How the test worked

Any investment manager with more than $100 million in U.S.-listed stocks files a Form 13F with the SEC every quarter. It lists what the fund held on the last day of that quarter.

The filing deadline is 45 days after the quarter ends. A stock a fund bought in the first week of January can stay out of public view until mid-May, and by then the fund may have bought more, trimmed it or sold it entirely.

Nobody outside the fund can buy at the fund’s price, so I built the test around the first day a follower could act:

1. For each fund, I took its most recent two to four quarterly filings in our data, and from each one the stocks it opened or added to. I used the largest moves in each filing, up to 25 per filing, so small housekeeping trades did not dilute the results.
2. Each pick was bought at the first weekly closing price after the 13F was filed.
3. Each pick was held to the latest weekly price in the data, the week of October 2, 2026.
4. For every pick, the same amount went into the S&P 500 (through SPY) on the same day and was held for the same time.
5. Every pick counts equally, whatever its weight in the fund. Filings less than 30 days old were left out because they have not had time to play out.

A fund “beats the index” here when the average return of its copied picks is higher than the average return of the S&P 500 over the same days. In total, the test covers 607 picks across the 28 funds.

The results

Coatue Management, Philippe Laffont’s fund, led the table. Its 24 copied picks returned 12.3% on average while the S&P 500 returned 7.7% over the same days, a lead of 4.5 percentage points. Li Lu’s Himalaya Capital came second, 0.8 points ahead, but on only 3 picks, which is too few to read much into.

The other 26 funds fell behind. Here is the full ranking, by how many percentage points each fund’s average pick was ahead of or behind the S&P 500:

1. Coatue Management (Philippe Laffont): +4.5
2. Himalaya Capital (Li Lu): +0.8
3. Point72 (Steve Cohen): -0.7
4. Renaissance Technologies: -0.9
5. Duquesne Family Office (Stanley Druckenmiller): -1.6
6. Starboard Value (Jeff Smith): -1.9
7. Bridgewater Associates: -3.3
8. Dalal Street (Mohnish Pabrai): -3.5
9. Appaloosa (David Tepper): -4.6
10. Viking Global (Andreas Halvorsen): -6.7
11. Scion Asset Management (Michael Burry): -6.8
12. Baupost Group (Seth Klarman): -6.8
13. Third Point (Dan Loeb): -7.9
14. Berkshire Hathaway (Warren Buffett): -8.4
15. Greenlight Capital (David Einhorn): -8.5
16. Tiger Global (Chase Coleman): -9.0
17. Fundsmith (Terry Smith): -9.7
18. ValueAct Capital (Mason Morfit): -10.5
19. Lone Pine Capital: -11.4
20. Markel Group (Tom Gayner): -11.5
21. Soros Fund Management: -13.4
22. Oaktree Capital (Howard Marks): -14.2
23. Pershing Square (Bill Ackman): -18.4
24. Ruane, Cunniff & Goldfarb (Sequoia Fund): -18.8
25. Trian Fund Management (Nelson Peltz): -22.8
26. Elliott Investment Management: -22.9
27. Akre Capital Management: -25.7
28. Carl Icahn: -28.2

Point72, Renaissance, Duquesne and Starboard finished within two points of the index, close enough to call a draw. Below them the gap widens quickly, and eleven funds trailed by more than 10 points.

Single picks tell the same story as the fund averages. Of the 607 copied picks, 212 beat the S&P 500 over their own holding period, about one in three (34.9%). Pick one of these trades at random and the odds were nearly two to one that the index would have done better.

Big gains that still lost

The ranking hides something that surprised me: several of the funds that “lost” made a lot of money.

David Einhorn’s Greenlight Capital is the clearest case. Its 23 copied picks come from two filings in late 2023 and early 2024, so they have been held for close to three years. They were up 54.7% on average. Anyone who copied them would have been pleased, until they compared that with the S&P 500, which rose 63.1% over the same stretches. The picks from the February 2024 filing did beat the index (61.4% against 53.9%), but the earlier batch, up 45.8% while the index rose 75.1%, pulled the average below it.

Michael Burry’s Scion is the same story on a smaller scale. Its copied picks rose 15.5% on average while the S&P 500 rose 22.4% over the same days. Many of Scion’s picks came from filings in 2025 and were held through a strong market, so the bar they had to clear was high.

That is why I care more about the benchmark than the headline return. “I made 15% copying Burry” sounds like a success story. Measured against the simplest alternative available, it left 6.8 points on the table.

Four funds, four lessons

Coatue: the one that worked

Coatue’s copied picks returned 12.3% on average, and 12 of its 24 picks beat the index on their own.

Most of the edge came from one filing. The 13 picks from its third-quarter 2025 report, filed in November 2025, averaged 21.7% against 14.8% for the index. One stock did much of the work: Marvell Technology (MRVL), up 214% from the copy date. The same filing also held the worst pick, Figma (FIG), down 42%. Copying Coatue worked because its style happened to suit the market in that window, and nobody could have known that in advance.

Berkshire Hathaway: a great investor makes a poor template

Warren Buffett’s 20 copied picks returned 0.4% on average. The S&P 500 returned 8.8% over the same days, and only 5 of the 20 picks beat it. The seven picks from Berkshire’s February 2026 filing averaged 0.8% while the index rose 13.1%.

Berkshire buys businesses it expects to own for many years. The 13F shows the position. It does not show the price Buffett paid, how much more he plans to buy, or how long he is willing to wait. A follower who buys months later at a different price, and checks the result a year after that, is running a different strategy with Buffett’s name on it.

Pershing Square: few picks, big swings

Bill Ackman runs a small number of large positions, so his fund had only 7 copied picks in the test. They averaged -0.8% while the S&P 500 rose 17.7% over the same days, 18.4 points behind. With so few picks, one or two positions decide the whole result. In the August 2025 filing, Alphabet (GOOGL) rose 71% after the copy date, but Hertz (HTZ) fell 69%, and those four picks averaged -6.7% while the index rose 21.1%.

Carl Icahn: last place

Carl Icahn’s 9 copied picks averaged -17.6% while the index rose 10.6%. Not one of the 9 beat the S&P 500.

Part of the reason is what his “new buys” were. Icahn Enterprises (IEP), his own holding company, shows up among his picks in two of the three filings, down 13% and 6% from the copy dates. A 13F records that like any other purchase, but buying more of your own company is a very different decision from buying a stock you think the market has mispriced.

Why copying 13Fs mostly fails

The managers on this list include some of the most successful investors of the past few decades. What fails is the copy, and there are four reasons.

The information is late. A 13F is a snapshot of the last day of a quarter, filed up to 45 days later. A trade made early in the quarter can be more than four months old when you see it, and much of the move that made the fund buy may already be in the price.

It is a snapshot. You see what the fund held on one day, not what it did in between. A position that appears as “new” may already have been cut by the time the filing is public.

It leaves things out. A 13F covers long positions in U.S.-listed stocks and some options. It does not show short positions, most foreign holdings, bonds or cash. A stock that looks like a bold bet may be one side of a hedge you cannot see.

It hides the price and the plan. The filing shows how many shares a fund held, but not what it paid or what it intends to do next. You will almost always pay a different price, and you will not know when the fund plans to sell.

So copying a 13F means buying a stale, partial picture of someone else’s strategy at a worse price, without knowing their reasons or their exit.

Three questions before you copy a fund

If a fund’s new position still interests you, I would ask three things first.

1. How much has the stock moved since the end of the quarter? If it is already up 20% since the date the 13F describes, you are buying a different opportunity from the one the fund bought.
2. Is this a real decision or a small change? A fund adding 2% to a position it has held for ten years sends a much weaker signal than a brand-new position that makes up 5% of the portfolio.
3. What would make you sell? The fund has its own exit plan, and you will not see it until the next filing, months later. You need your own.

How I use 13Fs now

I still read 13F filings every quarter. I use them as a list of ideas to research, and I stopped treating them as a list of stocks to buy.

When a careful investor opens a large new position, it is worth asking why, and then doing the work yourself at today’s price. Several quarters in a row show what a manager keeps adding to and what they give up on, which says more about their thinking than any single filing. When many funds pile into the same stock, that tells you the trade is crowded. And when someone online claims a famous investor is “all in” on a stock, the 13F lets you check.

For anyone who wants market returns without doing research, an S&P 500 index fund beat 26 of these 28 copy strategies in this test.

The full scorecard, with every fund’s picks quarter by quarter, is free at wallstpal.com/funds/scorecard. The data comes from SEC 13F filings and is rebuilt every day.

This article is for information only and is not investment advice. Past results of any strategy do not predict future results.

📩 Join Investor’s Handbook — get the Tuesday Digest, Thursday Weekly Framework, and free financial tools designed to help you make better decisions about investing, money, and building wealth.

Learn the framework. Run the model. Make the decision.

🎥 Watch Investor’s Handbook on YouTube · 𝕏 Follow Investor’s Handbook on X

I Copied the New Stock Picks of 28 Famous Hedge Funds. Only 2 Beat the S&P 500. was originally published in Investor’s Handbook on Medium, where people are continuing the conversation by highlighting and responding to this story.

Pricing a vanilla European option is easy. Pricing everything else is where quants lose their weekends.

Asian options need observation dates and historical fixings. Barrier products need to know whether the price ever touched a level. Add stochastic volatility or multiple assets, and you’re suddenly carrying extra state variables through every step of your model.

For these use cases, DolphinDB provides two general-purpose pricing toolkits that let developers focus on financial models and product logic rather than implementing numerical solvers from scratch:

• Monte Carlo: Simulate stochastic paths, define custom payoff logic, and obtain statistical estimates with standard errors and confidence intervals.
• PDE solvers: Solve 1D and 2D linear parabolic PDEs to generate full price curves or surfaces, with interpolated values and Greeks.

This article walks through both toolkits with practical examples, compares their performance against Python, QuantLib, and PriceLib, and discusses how to choose and validate a pricing method in real-world applications.

In this post

1. Monte Carlo: build path-dependent pricing models flexibly
2. pdeSolver: one solve, the whole price surface
3. Performance benchmarks
4. How to choose and validate a pricing method in practice
5. What this unlocks, and where we’re heading

1. Monte Carlo: Build Path-Dependent Pricing Models Flexibly

For Asian and barrier options, the payoff depends not just on the terminal price but on how the underlying got there. To price them, you simulate the asset’s random paths, compute the payoff under the product terms, and estimate the price statistically.

Writing that pipeline from scratch every time means re-implementing random number generation, correlation handling, batch computation and statistical aggregation again and again. But look closely, and products differ in just two ways:

1. How the underlying state evolves
2. When, and how much, the product pays

Path simulation and statistical estimation can be reused. DolphinDB’s Monte Carlo toolkit is designed around exactly this idea: you define state evolution and payoff logic as custom functions, and the toolkit orchestrates everything else.

Core capabilities

Three functions do the heavy lifting:

• mcSimulate: simulates stochastic processes and generates paths.
• monteCarlo: combines your evolution function, payoff function and simulation settings into a complete Monte Carlo pricing run.
• mcEstimate: produces statistical estimates from simulation output.

Two workflows cover most needs:

• Research and debugging: generate paths with mcSimulate, inspect them, apply custom processing, then estimate with mcEstimate. It's ideal for checking intermediate results, though storing full paths can consume significant memory.
• Production pricing: call monteCarlo directly. It advances the simulation in batches and aggregates results internally, which suits high-volume numerical estimation with a fixed payoff.

The core product-specific logic is captured in two functions: evolve and payoff. European, Asian, barrier, lookback and multi-asset options can all be defined on top of the same simulation framework. For stochastic volatility models, simply include volatility as an extra state variable in the simulation.

Example: an Asian option under GBM

Assume the underlying follows geometric Brownian motion (GBM). First, define how the price evolves:

def gbmEvolve(t, state, dt, noise, p) {
    drift = (p["r"] - p["q"] - 0.5 * p["sigma"] * p["sigma"]) * dt
    return state * exp(drift + p["sigma"] * sqrt(dt) * noise)
}

Next, define the payoff. This example prices both the arithmetic-average and geometric-average Asian call in a single pass:

def asianPayoff(paths, timeGrid, p) {
    n = rows(paths[0])
    m = size(paths) - 1
    prices = matrix(DOUBLE, n, m)
    for (i in 1..m) {
        prices[, i - 1] = paths[i][, 0]
    }
    arithmeticAverage = rowAvg(prices)
    geometricAverage = exp(rowAvg(log(prices)))
    T = timeGrid[size(timeGrid) - 1] - timeGrid[0]
    df = exp(-p["r"] * T)
    return df * matrix(max(arithmeticAverage - p["K"], 0.0),
                       max(geometricAverage - p["K"], 0.0))
}

Then run the pricing with the initial price, time grid, number of paths and model parameters:

res = monteCarlo(gbmEvolve, asianPayoff, 100.0,
    0..12 / 12.0, 100000, 1, p, opt)

gbmEvolve and asianPayoff carry your model logic. The remaining arguments set the initial state, simulation horizon, scale and configuration. The toolkit handles simulation and estimation, and returns the price estimate along with its standard error.

The takeaway: when product terms change, you only rewrite evolve and payoff. The simulation and estimation machinery stays untouched.

2. pdeSolver: One Solve, the Whole Price Surface

For European options, analytical solutions are available under standard assumptions. American options are more challenging because their value depends not only on the underlying price and time, but also on the optimal exercise decision at each point before maturity.

This makes PDE methods particularly useful: instead of computing a single price, they solve the pricing equation across a discretized space-time grid, producing a price curve or surface. The same framework can accommodate early-exercise constraints, different boundary conditions, time-dependent parameters, and more complex payoff structures.

Core capabilities

DolphinDB’s pdeSolver family solves 1D and 2D linear parabolic PDEs, with support for:

• Flexible PDE coefficients: configure diffusion, convection, reaction and source terms as constants or as functions of price and time (e.g., local volatility).
• Grids and boundaries: uniform grids and non-uniform sinh/tanh/log grids (refine around the strike or barrier), plus Dirichlet, Neumann and Robin boundary conditions.
• Time-Stepping Schemes: explicit, implicit Euler, Crank–Nicolson and Theta schemes, with Rannacher smoothing to suppress oscillations on discontinuous payoffs.
• Early-Exercise Constraints: American options are expressed through an obstacle configuration, with both PSOR and Penalty solution paths.
• 2D problems: ADI methods (Douglas, CS, MCS, HV) for models with cross-derivative terms, such as Heston, Basket and Exchange options.
• Pricing Results & Diagnostics: prices, interpolated values at specified points, Delta, Gamma and other Greeks, plus convergence and iteration diagnostics.

You describe the financial model by configuring the equation, terminal condition, boundary conditions and numerical parameters. You never implement the PDE algorithm itself.

Example: an American put

Spot 100, strike 100, risk-free rate 5%, volatility 20%, one year to expiry. Start with the equation coefficients, payoff and boundaries:

coeff = pdeCoeff1D(
    diff={S,tau -> 0.5 * volatility * volatility * S * S}, // diffusion
    conv={S,tau -> rate * S},                              // convection
    react=-rate                                            // reaction
)
payoff = {S -> max(strike-S, 0.0)}                         // payoff at maturity
boundary = [strike, 0.0]                                   // boundary values at both ends

The early-exercise constraint goes in the solver configuration through obstacle. Here we use PSOR:

method = dict(STRING, ANY)
method["obstacle"] = payoff
method["obstacleSolver"] = "PSOR"
method["psorOmega"] = 1.2
method["psorTolerance"] = 1e-10
method["psorMaxIterations"] = 20000

Finally, build the spatial and time grids and pass everything to pdeSolver:

spotGrid = pdeGrid1D(0.0, 300.0, 401, "sinh", strike, 3.0)
tGrid = (0..400) / 400.0
setting = dict(STRING, ANY)
setting["interpAt"] = 100.0
result = pdeSolver(
    coeff, payoff, pdeDomain1D(spotGrid), tGrid,
    boundary, method, setting
)
result["interpValue"]

The coefficients, terminal condition and boundaries define the pricing model; early exercise is a separate, pluggable constraint.

Product differences mostly come down to a few configuration choices:

• European: drop the early-exercise constraint.
• Barrier: encode the terms in the solution domain and boundary values.
• Local volatility: change how the diff coefficient is computed. The equation structure stays the same.

Grid construction and the solving workflow never need to be rewritten.

3. Performance Benchmarks

Speed matters well beyond a single valuation. It drives batch pricing, parameter sensitivity sweeps and complex product research. We benchmarked both toolkits against Python, QuantLib and PriceLib.

A note on methodology: libraries differ in model expression, algorithms and execution-level optimization, so these results compare end-to-end efficiency in specific scenarios and configurations, not absolute rankings. Where algorithms or precision settings differ, we say so explicitly.

Monte Carlo benchmark

All implementations ran 100,000 paths and 126 time steps, using a single worker and antithetic variates. Each test had one warm-up run followed by seven timed runs, and we report the median end-to-end time (DolphinDB timings include SDK round-trip overhead). All times are in milliseconds.

“Standard” and “log” are two ways of writing the state representation within the same engine. † QuantLib’s Heston uses its native FullTruncation variance discretization, which differs from the other implementations, so it is excluded from comparison. N/A means the interface is not supported.

What the numbers say

• vs. QuantLib: across the five comparable scenarios, DolphinDB’s standard formulation runs in roughly 1/18 to 1/77 of QuantLib’s end-to-end time, and the log formulation in roughly 1/8 to 1/145. Geometric Asian and discrete Up-and-Out with the log formulation are about 117x and 145x faster, respectively.
• vs. NumPy: DolphinDB is faster on European Call, Geometric Asian (log) and discrete Up-and-Out.
• Where it doesn’t win: on Heston Call, both DolphinDB formulations are slightly slower than NumPy, and on Arithmetic Asian the standard formulation is clearly faster than log. Execution efficiency depends on the state representation, the expressions and the implementation, so no single formulation is best everywhere.

For real workloads, evaluate numerical accuracy, algorithmic consistency and your target workload together.

pdeSolver benchmark

Tested on DolphinDB Server V3.00.6.4 (Linux), single CPU core, Release build. Each scenario uses its own spatial grid, time horizon and numerical scheme. Times are post-warm-up medians, in milliseconds.

What the numbers say

• vs. Python: DolphinDB is faster in all nine scenarios, up to 42.78x, with clear gains on Heston, digital and barrier options.
• vs. QuantLib: results depend on the scenario. DolphinDB is faster on digital, barrier, Local Vol and Quanto cases, while QuantLib is faster on BS European Call and Heston. The BS American Put compares different numerical algorithms, so we don’t treat it as a head-to-head performance result.

4. Choosing and Validating a Pricing Method in Practice

Monte Carlo and PDE each have their sweet spot, and performance shifts with model and implementation. In real projects, choose a method on three criteria, in order: model fit, numerical reliability, computational cost.

Step 1: Let the product’s characteristics pick the tool

First check whether a dedicated pricing function already covers your need. If not, decide between general Monte Carlo and PDE based on model state, payoff structure and the outputs you need.

These are guidelines, not hard rules. The final call depends on model assumptions, solver support, accuracy requirements and compute resources.

Step 2: Validate with numerical diagnostics

A price isn’t ready for production until it’s been validated, and different methods fail in different ways.

• Monte Carlo: check standard errors, confidence intervals and estimates across different simulation sizes to confirm statistical error is acceptable.
• PDE: refine the grid, adjust solver tolerances and check convergence of both prices and Greeks.

Where possible, cross-check against analytical solutions or an independent reference implementation. For complex products without closed forms, build a validation plan from statistical error, grid convergence and independent implementations.

Step 3: Evaluate performance once accuracy is met

Optimize on top of a correct, reliable model, not before. For Monte Carlo, tune path count, time steps and batch configuration. For PDE, test grid size, solver algorithm and execution path.

As Section 3 shows, performance varies by model and implementation. Before deployment, benchmark representative products and parameter sets, and keep model definitions, accuracy targets and test scale comparable.

In short: pick the method, validate the numbers, then measure speed. DolphinDB’s general Monte Carlo and PDE toolkits give you a reusable computation framework for that process, so you can spend your time on financial models and product terms.

5. What This Unlocks, and Where We’re Heading

Beyond raw numerical power, these toolkits give you a reusable foundation for developing, optimizing and integrating derivatives pricing models.

Flexible model building. Standard products use dedicated pricing functions directly. For exotics, path-dependent products and custom payoffs, you define state evolution and payoff logic, or configure PDE coefficients, boundaries, grids and solvers. No need to build a full computation framework for every new product.

Model development meets high-performance computing. Batch computation, vectorization and JIT support help optimize execution. Monte Carlo can be tuned through path count, time steps, batching and memory limits; PDE through grid size, precomputation and solver choice. Actual performance depends on model, algorithm, hardware and parameters, so validate it with comparable tests.

Pricing and business data in one place. Because pricing runs alongside DolphinDB’s data processing and analytics, market data, model parameters and results don’t have to shuttle between separate systems. You can build standardized workflows that manage product parameters, model configuration, numerical methods and result validation together, laying the groundwork for batch valuation, scenario analysis and downstream risk work, with less duplicated development and maintenance.

Built to keep growing. Directions we’re exploring include 3D PDEs, finite volume (FVM) and finite element (FEM) methods, Sobol sequences and quasi-Monte Carlo (QMC), control variates, Brownian bridge corrections for continuous barriers, optimal stopping algorithms for American and Bermudan options, model calibration, and automatic differentiation for Greeks. These are potential directions only; actual support will follow product progress.

Whichever numerical method you choose, the point is the same: reuse proven computation workflows and focus your effort on the financial model, the product terms and the business logic, not on re-implementing solvers.

Thanks for reading! Stay connected with DolphinDB by following us on X (@DolphinDB_Inc) and LinkedIn for the latest updates, technical articles, and product news. For more information, visit our website.

📩 Join Investor’s Handbook — get the Tuesday Digest, Thursday Weekly Framework, and free financial tools designed to help you make better decisions about investing, money, and building wealth.

Learn the framework. Run the model. Make the decision.

🎥 Watch Investor’s Handbook on YouTube · 𝕏 Follow Investor’s Handbook on X

Building Flexible Derivatives Pricing Models with Monte Carlo and PDE Solvers was originally published in Investor’s Handbook on Medium, where people are continuing the conversation by highlighting and responding to this story.