I didn’t know the final numbers, obviously, but the investment thesis, commentary, and general conclusions were usually inferable in advance. After the earnings report, I’d update my financial model and add color to anything notable before publishing.

Sometimes I’d spend less than an hour editing a report that looked like it had taken days to write.

Other analysts I worked with did the same thing. Our institutional clients expected research within hours of a company’s earnings announcement, and there wasn’t time to write a 15-page report from scratch.

But there was another reason this worked. After covering the same companies for several quarters, I generally knew what management was going to say. The earnings calls rarely changed my investment thesis. Most of the research had already happened.

Now that I use AI for my own investment research, I’ve been thinking about how much of that process individual investors can replicate. Particularly the preparation that happens before a company reports earnings.

What actually happens on an earnings call

If you’ve ever listened to an earnings call, you know how formulaic they can be. Management discusses quarterly results, highlights the areas performing well, acknowledges a few challenges, and reassures investors about the long-term outlook. It’s pretty predictable.

That’s because earnings calls are carefully prepared. Management has worked with investor relations and legal teams to decide what information to disclose and how to present it.

One thing I learned was to pay attention to the metrics management stopped discussing. For example, if a company had spent the previous four quarters highlighting the growth of a particular product and suddenly stopped disclosing the number, I’d want to know why. That omission might matter more than anything management chose to emphasize.

The Q&A is more useful than the presentation

The most interesting part of an earnings call is usually the Q&A, when analysts have an opportunity to challenge management’s assumptions.

A 2011 study published in The Accounting Review examined more than 10,000 earnings call transcripts. Researchers found that both prepared remarks and Q&A sessions conveyed information beyond what was already in the earnings release, but the Q&A was generally more informative.

That’s consistent with my experience. During prepared remarks, management gets to choose the topics. During Q&A, analysts can raise issues management might prefer not to discuss.

A 2026 study in the Review of Accounting Studies examined earnings calls from 2002 through 2017 and documented how executives redirect answers away from analysts’ questions.

Interestingly, this behavior wasn’t necessarily an indication of bad news. Management was more likely to change topics when business performance was strong, among other circumstances. So a non-answer isn’t necessarily a red flag, but it’s certainly something to pay attention to.

For example, imagine an analyst asks whether a company’s largest customers are reducing orders. Management responds by discussing the company’s strong pipeline of new business opportunities. That’s the kind of thing I’d flag for further research.

What our institutional clients actually paid for

I covered healthcare companies, primarily in diagnostics, laboratory services, and genomics.

Our clients were institutional investors. They had their own financial models, analysts, and investment opinions. They weren’t interested in paying us to summarize quarterly results they could read themselves. They wanted proprietary research that might help them anticipate business performance before everyone else. That led to some strange assignments.

I spent days calling STD clinics across the country. One of the companies I covered manufactured molecular diagnostic tests. I called clinics to understand testing volumes, available tests, and which diagnostic equipment they used. We used this information to develop independent estimates of demand.

I got my blood drawn three times in one day. I covered two of the largest laboratory testing companies in the country. A Silicon Valley startup was promising to revolutionize blood testing, and investors were concerned about the competitive threat. I visited three laboratories in one afternoon, documenting differences in the test results. (That startup was Theranos.)

I crawled underneath a DNA sequencing machine to read its serial number. We wanted to estimate how many sequencing machines a company had sold. During a laboratory visit, I got down on the floor to find the serial number on a new machine, hoping to estimate the company’s sales that quarter.

This is what some of the research behind those polished Wall Street reports actually looked like. A junior analyst crawling around a laboratory trying to read a serial number so a hedge fund could refine its revenue estimates.

The point of all this work was to develop an independent view of business performance. If we could estimate a company’s quarterly revenue more accurately than other investors, that information might be valuable. But even that wasn’t enough.

We needed to understand whether our expectations differed from those already reflected in the stock price.

Good earnings don’t necessarily mean a good investment

This is one of the most important things for individual investors to understand.

A company can report an excellent quarter and still see its stock decline.

Suppose a company reports 20% revenue growth and beats Wall Street’s earnings estimates. But what if investors were expecting 30% revenue growth? Or the stock is trading at a valuation that assumes the company can maintain 30% growth for the next five years?

Suddenly, the results don’t look so impressive.

The opposite can also happen. A company can report declining earnings and see its stock rise because investors had expected an even worse result.

Analyst consensus estimates are a useful reference point, but they don’t necessarily capture everything investors have priced in. Valuation matters, too.

This is why I don’t think investors should pay too much attention to the initial market reaction after an earnings announcement.

The questions I’d rather answer are whether the company’s long-term prospects have improved or deteriorated, and whether the current valuation reflects those changes.

How I’d use AI to analyze earnings today

AI has made much of the research preparation I used to do on Wall Street available to individual investors.

You don’t need access to an expensive Bloomberg Terminal to review historical financial results. Companies publish their quarterly and annual reports through the SEC’s free EDGAR database.

Most also make earnings releases, presentations, and call recordings available on their investor relations websites.

Here’s the process I’d use.

Before earnings: establish your expectations.

I’d start by collecting the company’s recent financial filings, earnings releases, and several quarters of call transcripts.

Then I’d have AI compare management’s previous guidance with actual results and identify the operating metrics that matter most.

For example, I’d ask:

Review the last four quarters of this company’s earnings materials. Identify the three to five operating metrics most important to its long-term value. Track how those metrics have changed, compare management’s guidance with actual results, and identify any metrics the company has stopped disclosing. Cite the original documents for every finding.

This gives you a starting point before management presents the latest quarter.

I’d also establish what I expect to happen and what evidence would make me reconsider my investment thesis.

Otherwise, it’s too easy to rationalize disappointing results after the fact.

After earnings: focus on what changed.

Once the company reports, I’d upload the new earnings release and call transcript alongside the prior-quarter materials.

Rather than asking AI for a general summary, I’d ask it to compare the new information with our previous expectations.

Did revenue growth accelerate or decelerate? Did management change its guidance? Are margins improving for sustainable reasons? Has management’s explanation of its competitive position changed?

I’d pay particular attention to how management answered questions about the business’s biggest risks.

AI is useful here because it can compare several quarters of commentary without having to manually search through each transcript.

You can ask it to identify every instance where management discussed a particular competitor, customer group, product, or financial metric.

You can also ask it to find questions that management didn’t directly answer.

I would verify the underlying passages myself. A model might correctly identify a change in language while incorrectly interpreting its significance.

Finally: update the investment thesis, not just the quarterly forecast.

Suppose a company misses quarterly revenue expectations by 3%.

If the reason is a delayed customer order that will be recognized next quarter, the long-term impact could be negligible.

If the reason is a competitor introducing a superior product, that same 3% miss could be an early indication of something much more serious.

The financial result is similar. The investment implications are completely different.

This is where I think an investor’s own judgment matters most.

AI can help identify and investigate the differences. But I still want to decide for myself whether the new evidence changes the company’s long-term outlook.

You don’t have to compete with Wall Street every quarter

One thing I appreciate more now that I’m investing my own money is that I don’t have to predict next quarter’s results.

When I was in equity research, anticipating quarterly performance was an important part of the job. Institutional clients wanted information that could help them position ahead of earnings.

As an individual investor, I have a different objective. I’m looking for businesses that can become materially more valuable over the next five years. I don’t necessarily need to know whether they’ll beat earnings estimates next Tuesday.

I do need to know whether their competitive advantages are holding up, whether the growth opportunity is still attractive, and whether I’m paying a reasonable price.

Quarterly earnings are useful checkpoints to test whether the business is progressing as expected. They aren’t necessarily reasons to buy or sell.

AI makes those checkpoints much easier to manage. I can follow more companies, review more historical information, and investigate changes in their businesses without spending days reading transcripts.

But the preparation is still important. When I worked on Wall Street, I could write most of my earnings report in advance because I already understood the business and knew what I was looking for.

I think individual investors should approach earnings the same way. Know what matters before the company reports. Use AI to find what changed. Then decide whether the change actually matters to your investment.

I previously wrote about how your algorithm is your portfolio. Your social media feed has too much control over where you attention goes, so prune your feed carefully.

Your social media feed is also planting investment ideas in your head.

You see a ticker mentioned on X, Substack, and Reddit… and next thing you know, you have ChatGPT or Claude running a comprehensive investment analysis. And so is everyone else.

How stocks end up on your watchlist

Most individual investors don’t have a formal process for finding companies to research. They tend to add stocks to their list to research as they hear about them.

Once social media enters that equation, it introduces a certain selection bias.

Recommendation algorithms decide which content gets distributed based largely on what people engage with. And the AI models running you investment analyses are trained of this data.

The stocks getting the most attention are the most controversial ones. Not necessarily the ones that present the best investment opportunity.

FINRA’s 2024 investor survey found that 26% of investors use recommendations from social media influencers when making investment decisions. Among investors under 35, that figure rises to 61%. Stunning, and, the death of traditional equity research.

There are thousands of publicly traded companies. How many of them have you never considered simply because they never came up in your social media feed?

AI is remarkably good at researching popular stocks

I recently wrote about a study testing how AI models manage stock portfolios. Researchers tracked the stock selections of ChatGPT, Claude, Gemini, and Grok.

The resulting portfolios had about 41% allocated to semiconductor stocks, compared with roughly 20% in the S&P 500. The models tend to favor companies dominating the headlines. They constructed portfolios with significantly more risk than the general market.

Keep in mind that the study covered just eight months during a particularly strong market. It’s not proof that AI will always favor the most popular stocks.

But the results are worth noting.

We have tools capable of analyzing enormous amounts of financial information, yet their recommendations can end up looking suspiciously similar to whatever investors are already excited about.

There’s a reason for this. AI models learn from information that already exists. Popular companies have an abundance of articles, analyst reports, interviews, and online discussions available for the models to draw from.

A company with limited analyst coverage may have excellent financial statements but relatively little written about it. Unless the research process deliberately searches for these businesses, it’s easy for them to be overlooked.

That strikes me as a missed opportunity.

Finding the next investment

In my current investment research process, I look for companies with durable competitive advantages, a long runway for growth, and valuations that don’t already assume everything will go perfectly.

I’m particularly interested in businesses that have room to become materially larger over the next five years.

I’ve been using a stock screener to make a short list of companies that have the characteristics I’m interested in, and using different AIs to debate the merits of each. You can learn more about my current process here.

Popularity isn’t an investment thesis

I made 34x on Applied Digital after holding an unpopular investment thesis for about a year and a half. The market had largely written off the company, but I saw potential.

I’ve also invested in emerging technologies that later became wildly popular. In 2022, I bought publicly traded quantum computing companies with the intention of holding them for 20 years. By 2025, some of those stocks had appreciated around 50x. I sold because the valuations had gotten far ahead of the underlying businesses.

Being early matters, but so does recognizing when the opportunity has already been priced in.

I’m not opposed to investing in popular companies. Nvidia is an obvious example of a stock that rewarded investors despite years of attention and increasingly high expectations. Popularity alone tells you little about whether a stock is attractive.

The issue is how we decide which businesses deserve our attention.

AI has made it easier than ever to research practically any company. I think individual investors should take advantage of that by expanding their search beyond the stocks they already know.

Otherwise, we’re using extraordinary technology to research the same handful of companies that the internet was going to tell us about anyway.

The workflow is designed to separate the tasks AI is uniquely good at vs. what humans are uniquely good at. AI is excellent at gathering evidence, comparing sources, modeling scenarios, and pressure-testing assumptions. The investor is best suited to retain responsibility for judgement.

A core design principle is that consistency matters more than having a perfect prompt. Once calibrated, I use the same process and definitions across companies so the outputs are comparable. Even slight changes to prompts will produce vastly different outputs. Running each investment through the same process produces more comparable insights.

What I’m Looking For

I’m looking for companies that have room to become materially larger, have durable competitive advantages, and do not already have unachievable expectations priced in.

Keep in mind that there’s no such thing as a perfect investment. There are only those that fit your risk tolerance, time horizon, and goals better or worse. Advice from someone who doesn’t know your situation is close to worthless. I’m sharing this as an illustration of what works for my portfolio, not as a recommendation of how yours should look.

My workflow assesses business quality before it considers valuation. This is intentional. Determine which companies deserve a premium in the first place. Then consider what premium is fair.

The Workflow

The workflow follows these steps:

1. Primary Analysis

   1. Model: GPT-5.6 Sol, High reasoning + current web search turned on

   2. Output: Business quality, runway, reverse valuation, Bear/Base/Bull, risks

2. Independent Red Team

   1. Model: Claude Opus 5.5, fresh context + web access

   2. Output: Challenges assumptions, comparables, correlated risks, terminal valuation

3. Reconciliation

   1. Model: GPT-5.6 Sol, High reasoning

   2. Output: Adjudicated final targets, CAGR, permanent-loss risk, portfolio fit

4. Tracking

   1. Tool: Master spreadsheet + individual stock tabs

   2. Comparable opportunity set across the full research universe

5. Finalist Diligence

   1. Model: GPT-5.6 Sol Deep Research

   2. Extra diligence for top candidates or unresolved complex cases

How to Use the Outputs

The master spreadsheet is used as a comparison tool. I organize company-specific research on to individual tabs. The key metrics ladder up to an executive summary view that compares opportunities.

I prefer using Google Sheets because it allows me to link to other tabs for faster navigation. Using Google Finance formulas keeps the spreadsheet up-to-date with real-time market updates.

The outputs from the AI models’ research and analysis give me a few comparable metrics I use to evaluate investments.

• Base, bear, and bull case targets.

  • Bear, base, and bull case targets are evaluated over a 5-year period. 1-year estimates tend to be heavily influenced by sentiment. A longer time frame is more difficult to predict reliably. Five years seems like a sweet spot.

• Portfolio fit categories.

  • Pass: Economics, valuation, expected return, or risk make this investment unattractive.

  • Watch: An otherwise attractive business or opportunity, but the current valuation or an unresolved issue makes it unattractive at this time.

  • Compounder: This indicates a high-quality business with credible upside and acceptable permanent-loss risk.

  • Asymmetric: Outcomes are binary. Large upside potential coupled with meaningful permanent loss risk.

• Permanent-loss Risk.

  • Scored on a scale of 1-10. The model is instructed not to conflate volatility with risk. Risk is defined as the probability of permanent loss. This is more appropriate for a longer-time-horizon investor.

• Required Expectations.

  • A current stock prices reveals many combinations of expectations around growth, margin, dilution, and multiples. The reverse valuation exercise asks what operating performance would be required for an investor to earn a 10% annualized return over five years. This is used as a common measuring stick.

The Role of the Human Investor

AI is a good tool to use for gathering information, revealing embedded assumptions, and pressure-testing estimates. But it should not be used to make judgement calls. The final capital allocation decisions are left to the human. Once the AI analysis is complete, evaluate for yourself which assumptions are believable. Only you can decide how much risk you are willing to tolerate. Don’t leave those judgement decisions to AI.

Imagine your AI is like an intern. A really smart intern: fast, well-read, and eager to help. But also new to the job and have a “fake it ‘til you make it” kind of attitude. It doesn’t tell you when it doesn’t know an answer. It just makes up an answer that sounds plausible.

Research shows that many of AIs failure points are predictable. More on that below.

It’s Predicting, Not Remembering

Every AI model has a training cutoff. This is the date when it last read anything. After that, its training data stops.

Researchers Lopez-Lira, Tang, and Zhu found that when you ask a model about anything before its cutoff date, the model remembers an answer based on its training data. It is not forecasting or predicting anything. This applies to stock prices, economic data, or earnings calls.

The researchers tried telling the model to ignore what it knows. They also tried hiding the company’s name. Neither of these approaches worked. When you ask the model about anything after the cutoff, the answers are rarely accurate.

Some AI tools connect to the live internet. ChatGPT, Perplexity, and Claude can search the web when you ask something current. That’s a separate tool layered on top of the model. It does not update the model’s underlying training data.

It Tells You What You Want to Hear

In a 2025 study, researchers had the user ask AI a question with a wrong opinion. The model agreed with the user, on average, 63.7% of the time. Various individual models agreed with the wrong opinion 46.6-95.1% of the time.

Research has also found that telling the model to act like an expert in the subject does not improve the result.

For tasks that involve math or coding, telling the AI to act as an expert can actually make its answers worse. No facts have been added to its training data that would impart expertise.

Ask It Three Times

Researchers tested the same finance questions with AI models 50 times for each of five tasks. The testing produced over 3.4 million answers in total. Simple tasks, like sorting a sentence into positive or negative, often returned the same result. Complicated tasks, like writing a summary or making a prediction, varied a lot from run to run. A more advanced model wasn’t necessarily a more consistent one.

Asking the same question three to five times and comparing the answers made the results more reliable.

The Portfolio a Chatbot Hands You

Researchers studied how an AI Chatbot would manage a stock portfolio. They tracked daily stock picks from ChatGPT 5.0 and 5.2, Claude Sonnet 4.5, Gemini 2.5 Flash, and Grok 4.1.

The AI portfolios put about 41% of their money into semiconductor stocks, compared to about 20% for the S&P 500. Their average stock carried a beta near 1.6, meaning it swings in either direction roughly 60% more than the overall market in either direction. When researchers told the models to keep the beta between 0.9 and 1.1, the models steered toward the top of the range and sometimes disregarded the instructions and exceeded it.

The models tended to pick whatever stocks were being discussed in headlines, and build concentrated positions in those names.

Keep in mind that this study covers eight months of an unusually strong stock market. Nobody has tested what these AI portfolios do once the market turns down.

The Checklist

AI is good at quickly ingesting a pile of numbers and telling you what they mean. If you give it a company's raw financial statements with the name stripped out, GPT-4 called the direction of future earnings right 60% of the time, beating the average human analyst's 53%.

When it comes to multi-step judgment calls, the AI answer often falls apart.

Ask these questions before you trust an answer from AI.

If you ask AI to pick stocks or construct a portfolio, it will likely suggest a concentrated bet on whatever’s trending. AI can be highly useful for stock research, if it’s used effectively.

After researching and experimenting with these models, I thought it would be helpful to put together a guide. The goal is to help you choose the best model for each task and to get the most out of the model.

Intelligence and Behavior

Like humans, there is no single model that is best at everything. Differences in models can be categorized into two buckets:

• Raw capability. Think of this like IQ. Some models are better at problem solving, reasoning, and comprehending large amounts of information.

• Behavioral characteristics. Some models are better at following instructions. They take initiative, speak up, and challenge assumptions. Some keep digging until they have exhausted a question, often going down a few rabbit holes along the way. Others will quickly answer and want to move on.

Models vs. Products Distinction

One of the first things to clarify is the distinction between models and products. People often use the terms interchangeably, which creates confusion.

• A model is the underlying intelligence. Example: GPT 5.6 Sol, Claude Opus 5, or Grok 4.6.

• The product is the application through which you interact with the model. Example: ChatGPT, Claude, Microsoft Copilot, Perplexity.

• The mode is the amount of reasoning compute. Example: low, medium, high, max.

A model might work differently or less effectively when it’s embedded in a different product.

For example, Perplexity and Microsoft Copilot are product wrappers through which users can access various models. They then add their own tools, document handling, and instructions around those models. For this reason, you can have a different experience using the same models depending on what product you access it through.

Even when you use Claude or ChatGPT in your browser, your experience is shaped by their product layer too.

This evaluation of models will not cover an evaluation of products. Note that the benchmark scores provided in the guide are based on a stripped-down setup.

Different Tools Do Different Jobs

All that said, my first recommendation is to not pick just one model and try to use it for everything.

In terms of raw intelligence, some models are clearly ahead of the pack. Models with higher cognitive ability are better at reasoning through multiple steps. They excel at understanding complex documents and identifying non-obvious relationships.

The AI landscape is constantly changing as companies release new models all the time.

Anthropic’s Claude Fable 5.1 and OpenAI’s GPT-6 Astra top most intelligence benchmarks. Opus 5 is a close runner-up.

Intelligence is only one dimension to consider. A model’s greatest strength can also be its greatest weakness. Super intelligent models are known for overcomplicating things, changing the scope of the ask, and wasting time (and credits) going down irrelevant rabbit holes.

The smartest model might not always be the best one for the job.

Most people over-index on a model’s raw intelligence and ignore its behavioral characteristics. This is a mistake. Behavior often matters more.

The most important behavioral failure is sycophancy. This is AI’s tendency to conform to whatever you already believe instead of what is true. It’s a widespread feature of AI. A 2026 study by Carnegie Mellon and Emory tested 17 models. It found that sycophancy is a common failure mode in all of them.

AI loves to tell you what you want to hear.

Finance Industry Benchmark

Before you assign any work to a model, you need to understand what it can actually do in this domain.

Researchers test AI models against a standardized benchmark: Vals AI’s Finance Agent v2.

This tests models using 927 expert-reviewed questions. These questions match the analytical depth expected of a second or third-year investment banking analyst.

Every model gets the same toolkit: SEC EDGAR search, web search, a page parser, a retrieval tool, a calculator, and price history. Two hours per task.

It is graded by weighted checks. Certain critical facts are marked as dealbreakers. If the model misses one, the answer scores zero regardless of what else is right.

Here is the best score achieved by any model in the entire field, broken out by type of task:

The best model on the market fails roughly two out of three valuation modeling tasks. And that’s the leader, not the average. Keep that in mind.

Let this ranking give you an operating rule. Delegate the tasks AI models are strong at (retrieval and summarization). Draft-and-verify the mediocre ones (reconciliation). Leave the modeling to yourself.

The Most Expensive Models Are Not Better

The most expensive models aren’t winning.

On Finance Agent v2, Claude Sonnet 5 (53.9%) beats GPT-5.6 Sol (53.8%), Grok 4.6 (53.7%), and GPT-6 Astra (53.5%), one of the priciest models on the market. Within OpenAI’s own lineup, the budget model GPT-5.6 Luna outscores flagship Sol. Vals: mid-tier models trail the leaders by a few points, at a fraction of the price.

Save yourself the money and don’t always reach for the most expensive model thinking it will perform better.

Models are Like Employees

Think of models as employees rather than as machines or calculators. Each has different strengths, weaknesses, and quirks. The magic happens when they are in the right seat where they can play to their unique strength.

That said, here’s how I would use different models to assemble your AI dream team.

The Lean Operating Model

To avoid stacking up AI subscriptions, here’s what to buy to cover most of this. (Figures as of September 2026, subject to change.)

ChatGPT Plus + Claude Pro: $40/month

ChatGPT Plus ($20) is the first individual tier where GPT-5.6 Sol appears in standard chat, at Medium and High reasoning. Sol Pro and Extra High require the $100 or $200 Pro tiers. GPT-6 Astra appears in regular chat only on Pro, Business, and Enterprise.

Claude Pro ($20) covers three seats: Opus 5 for filings, Sonnet 5 as your associate, and Opus 5 again as the red team.

Claude Fable 5.1 is not included in Pro. Don’t upgrade to Max ($100) just for the red team. Using Opus 5 with a depersonalized prompt is an adequate replacement.

Here’s what you lose with this leaner setup:

The Market & Earnings seat. Sol is a decent substitution. Although, you’re giving up native audio and video input.

Gemini’s financial expertise. The free tier of Gemini uses Gemini 3.6 Flash. It scores 56.3% on the finance benchmark, while 3.8 Flash scores 61.4%.

The real problem with Gemini 3.6 Flash is that it allows such a small context window, you can’t even fit a 10-K in it.

If I Had to Pick

If you can only have one model, the one I would recommend depends on your workflow. Use ChatGPT Plus if your work typically starts with a question and goes out to the web. Use Claude Pro if your work usually starts with a document.

Volatility doesn’t necessarily increase growth. It widens the possible range of outcomes. Compounding is geometric, not arithmetic. That’s why more volatility often leads to worse outcomes. So “high risk = high reward” is mostly a myth we tell ourselves.

A better way to get high growth is to add leverage to a lower-volatility portfolio.

The search for cheap leverage

I found very few borrowing options available to individual investors. Most of them did not make financial sense. To profit from leverage, you need to borrow at a rate that is substantially lower than the investment’s expected return. Analysts project the U.S. stock market will grow at 6.7%. Almost no one will lend to an individual investor meaningfully below that rate.

Most brokerage margin accounts run in the low double digits range. Even cheaper brokers still land in the mid-single digits depending on balance.

I checked low-interest personal loans. They explicitly restrict what the proceeds can be used for and don’t allow securities investing.

It took some digging, but I found a few viable options where leverage is attainable at a rate that makes sense.

Real estate comparison

I would argue that most of what makes real estate investing attractive is actually just leverage. No other asset class that retail investors can access offers the debt terms that real estate does. Without debt, real estate is a lot less accessible and a lot less attractive to retail investors.

Real estate also tends to be a less volatile asset than stocks. Huge drops in value are less common. Which makes the use of leverage less intimidating.

Housing is a less-frequently-priced asset. Stocks get repriced every day the stock market is open. But houses only get repriced when they are bought, sold, or appraised. The volatility still exists. You just don’t see it when the asset gets repriced that infrequently.

More importantly, your mortgage does not get marked to market. Your mortgage doesn’t get called in simply because your home’s value drops, as long as you keep making payments. When you apply debt to a stock portfolio, a decline in the value of the underlying assets can trigger a margin call directly.

Why use leverage at all

Top investors, like Warren Buffett, used leverage to achieve outstanding returns. Research on Berkshire Hathaway estimates Buffett ran the portfolio at roughly 1.6-to-1 leverage on average. That leverage came almost entirely from insurance float, not from margin debt. Float isn’t marked to market and can’t be called the way a margin loan can. It’s a safer form of leverage than what is available to an individual investor. Nevertheless, the underlying principle holds. Applying leverage to a lower-volatility, high-quality portfolio helped Buffett generate outsized returns.

Options for individual investors

Most low-interest loans specify that you can’t use the proceeds for speculative investing, and standard brokerage margins are typically priced high.

A couple other options exist.

ETFs with leverage built in. Funds like RSSB or NTSX hold a base of equities and Treasury futures to create capital-efficient leverage. The leverage lives inside the fund, so there is no margin call risk to the investor personally. RSSB launched in 2023, so performance data is limited. Worst case scenario, stocks and bonds move in the wrong direction at the same time. This happened in 2022 and NTSX declined -30%, compared to -24% for the S&P 500.

Box spread loans. This may be appropriate if you have a high brokerage balance and experience with trading options. Box spread loans typically reflect Treasury rates plus a small margin. This more closely reflects the rates institutional investors have access to. They are currently trending around 4-5%, even for long duration loans.

Setting up the box spread

I recently took out a box spread loan through my brokerage, Charles Schwab. To do it, I needed approval for the right level of options trading. I could only place the trade through their thinkorswim platform. I used ToS’ paperMoney simulator to practice placing the trade beforehand, and used AI to double check I had set it up correctly. This is something where you can absolutely get burned if you don’t know what you’re doing. Brokerages gatekeep advanced options trading strategies for a reason.

I used SyntheticFI to track rates before committing.

One thing worth knowing going in: there is a difference between American-style and European-style options.

• American-style options can be exercised any time before expiration, creating early assignment risk. Your position can get called away unexpectedly, as one guy learned the hard way.

• European-style options (including index options like SPX and XSP) only exercise at expiration, avoiding that risk entirely. You are paid upfront, and the trade settles in cash at expiration.

I built out a Google Sheets model to size and set up the trade. I used Claude to build a small interactive tool to model decline cushions against different loan sizes.

Once my order was filled and I received the cash, I invested it in ETFs that align with my investment goals.

When my loan comes due in September 2027, I can either cover it by selling assets or by taking out a new box spread loan. There is a rollover risk at this point. If both asset prices are down and interest rates are substantially higher, I could be forced to sell at a bad time or refinance at a much worse rate.

Since a box spread nets out as a gain or loss on options contracts, the “interest” is treated as a capital loss. That could be used to offset capital gains for tax purposes.

Sizing the loan: stress testing

Before committing to a size, I ran the numbers against how bad past market declines have actually been.

• 2008 -57%

• The dot com crash -49%

• COVID 2020 -34%

• 2022 -25%

I used this to model at what point a decline would trigger a maintenance margin call, given my portfolio size and maintenance requirement.

At my chosen loan size (~31% of my pre-loan portfolio value), the portfolio could decline roughly 66% before triggering a margin call. This would be beyond any of the historical drawdowns in the last 50 years, so, unlikely.

The cushion assumes that the maintenance requirement stays fixed. The brokerages set these requirements and adjust them at their discretion. They typically raise house maintenance requirements during periods of high volatility.

Where this leaves us

I’ll be tracking the performance of what I invested in with my box spread loan and see how well the returns compare to the cost of borrowing. Follow along to see how it does.

Adobe (ADBE)

Adobe makes industry-standard software for graphic design, photography, video editing, and document management. The company has over 850 million monthly active users. In a survey of ~16,000 creators, 86% said they already use generative AI. In filmmaking, 85% of the films premiering at the 2026 Sundance Film Festival used Adobe software. Adobe’s generative AI platform, Firefly, can generate image, video, and audio. 99% of Fortune 100 companies have used AI in an Adobe app.

Yet, Adobe stock is down ~24% in the past year. Adobe currently trades at ~10x projected earnings. This is a discount relative to the broader software sector.

Adobe’s annual revenue run-rate (“ARR”) of $26-27B is 9-13% of its total addressable market (“TAM”) of $200-300B. The market is not pricing in any share gains from AI for Adobe. The market is betting that rivals (Figma, Canva, OpenAI, Midjourney) capture the AI upside.

The Bear Case

Bears argue AI is eroding Adobe’s per seat subscription model, and that it’s happening faster than Adobe can monetize its own AI tools. It’s also happening without a permanent CEO or CFO in place. CEO Shantanu Narayen announced his exit in March 2026 and CFO Daniel Durn left in June 2026. Neither role has been filled.

Enterprise marketing budgets are under pressure. Customers are switching to AI-generated images. As such, Adobe’s stock photo licensing business is shrinking faster than management anticipated.

Adobe has embedded its creative and document tools inside Microsoft Copilot and ChatGPT. Bears argue this turns Adobe into a commodity plugin, rather than a main platform, giving them less pricing power.

A $150M DOJ settlement with Adobe forces them to simplify the customer cancellation process. This will likely result in increased future churn, reversing Adobe’s historically high customer retention rate.

The Bull Case

Rather than disrupting Adobe’s business, AI could be a growth accelerant. Adobe’s AI revenue is growing rapidly, but still only accounts for ~2% of revenue. Adobe continues to invest in AI, with recent acquisitions such as Topaz Labs. Customers are adding Adobe’s AI solutions on top of existing plans. Adobe’s tools are discoverable inside ChatGPT (900M+ weekly active users) and Microsoft 365 Copilot. This puts Adobe into the hands of a huge software audience without spending its own marketing budget.

Adobe has a few advantages that AI can’t easily replicate.

• Technical: Adobe is wired directly into enterprise’s customer records, online stores, and data. Ripping it out would be costly, high-risk, and time consuming. It is not as simple as a subscription cancellation.

• Legal: When Adobe’s AI makes a picture, Adobe knows exactly where the source material came from, because they only trained it on stuff they had permission to use. Some other AI tools may have secretly used artists’ work without asking, which means a company using them could get sued.

• Data: Enterprises store their brand libraries and campaign history inside Adobe’s products. Adobe can use that to build a custom AI just for that company.

Adobe can afford to be patient. Its cash reserves ($9B+ in free cash flow) and high profits margins (~89% gross margins) enable it to outlast its cash-strapped rivals.

Even if nothing else changes, a strong CEO hire could reset the multiple overnight. Especially a CEO that is well-versed in AI.

What to watch for

The Q3 earnings call will take place on September 10, 2026. Here’s a few things to listen for:

• Organic net new ARR: If it’s worse than -10% YoY, this confirms the bear case is playing out.

• Firefly/AI-first ARR: does it clear ~$400M and hold its ~50% quarter-over-quarter pace?

• Operating margin: does it keep compressing or stabilize?

Also watch for a new CEO to be named. It’s been about five months since the previous CEO announced he is stepping down. I would expect the stock to react well to a credible, external, AI-native hire. If Adobe picks an internal, business-as-usual candidate instead, that’s a bad sign. It probably means the board could not attract outside conviction.

I hear a lot of comparisons of AI to the internet and of the economics of AI to the dot-com bubble. I see corporate executives investing billions into building AI systems for their companies. At the same time, they are paying billions in severance to part ways with their human employees.

I started using AI pretty early on. I questioned ChatGPT with doubt when it first came out. My first prompt asked it to write a real estate listing for my house. I was amazed at how good it sounded, considering how little information I provided. I swore I would never need a realtor again.

Yet the more I used AI, the more errors I found it making. By the time I’d spent hundreds of hours with it, I hardly trusted it with simple tasks. It sounds confident but makes basic errors. These errors are hard to catch because they seem random. I learned to check every step before mistakes could compound. I became cynical. How can an executive trust AI to run a company? I can’t even trust it to find a good restaurant!

I’ve come to think there’s an S-curve to using AI. It mirrors Gartner’s Hype Cycle: Peak of Inflated Expectations → Trough of Disillusionment → Slope of Enlightenment → Plateau of Productivity. For the first 100 hours, you think it’s amazing and capable and eagerly hand it tasks. For the next 100 hours, you don’t trust it to do anything right. Eventually, you settle somewhere in the middle. You understand the difference between what it excels at and where its blind spots are.

As we examine the valuations of AI companies, we need to consider where the market stands in that cycle. Does the market think too much of AI, or not enough?

I dug into the numbers. My conclusion is that AI is a real economic transformation, and the market as a whole is not a bubble. Localized bubbles exist in certain companies and financing structures. Insiders even acknowledge this. Sam Altman has said, “someone is going to lose a phenomenal amount of money.” Some high-quality companies seem temporarily discounted due to AI fear. I believe this mispricing deserves its own focus.

How value gets priced

A company’s value is the net present value of its future cash flows. P/E multiples are shorthand for that equation:

AI boosts earnings in two ways:

• Directly through new products and better pricing

• Indirectly by reducing errors and cutting costs

A high multiple isn’t automatically irrational. It’s a claim about future growth and durability. The question is whether the company lives up to it.

Is there real demand?

The first indicator of a bubble is that there’s no demand. Flash back to the early 2000s, when companies like Pets.com and eToys imploded. These businesses burned through cash on customer acquisition but never found a path to profitable earnings. The companies we are examining today have established businesses and generate real earnings.

Analysts now expect the top five hyperscalers to spend $697 billion on AI infrastructure this year, with Goldman Sachs projecting $5.3 trillion by 2030. J.P. Morgan banker John Servidea called AI financing “the biggest secular theme in our professional lifetimes.”

The spending on AI infrastructure already exceeds the 1990s telecom and internet infrastructure build out. Hyperscaler AI capex is expected to reach ~3% of GDP by 2027, surpassing the telecom and internet buildout of the 1990s (1.0-1.2% of GDP). Capex doesn’t capture all of the spend. A typical 1-gigawatt AI data center costs about $38 billion to build and roughly $0.9 billion a year to run. Unlike capex, opex represents ongoing spend.

None of that investment matters unless revenue follows. So far, it is. OpenAI’s revenue grew from $2 billion in 2023 to $6 billion in 2024 to $20 billion in 2025. Anthropic’s revenue grew from a $9 billion run rate at the end of 2025 to reportedly $30 billion by April 2026. (For the record, OpenAI disputes this. They estimate Anthropic’s revenue is closer to $22 billion.)

Are the valuations reasonable

The hyperscalers themselves don’t appear overvalued.

If we compare Nvidia today to Cisco back in the dot-com era, Nvidia carries a much lower valuation multiple. Cisco traded at a trailing P/E ratio over 200x. Nvidia currently trades at 34x trailing P/E. Nvidia’s valuation is rich, to be sure, but it’s much more justifiable. Nvidia delivered $215 billion in annual revenue, up 65% y/y.

Many of the hyperscalers valuation multiples are lower today than they were a decade ago. Ten years ago, they were earlier in the growth curve. As more mature business today, requiring more reinvestment, investors are paying less for each dollar of earnings. So it makes sense that the multiples have come down. On this measure alone, it doesn’t look like a bubble.

Nvidia recently partnered with six asset managers to raise $500 billion for AI infrastructure buildout. Wall Street currently views data centers as expensive, risky, and unproven. The goal of the most recent deal is to de-risk them. Computing infrastructure is becoming a durable asset class with predictable cash flow, like commercial real estate.

Reasons to be skeptical

Michael Burry is a vocal bear. His argument is that hyperscalers inflate their earnings by depreciating hardware too slowly. Nvidia’s chips lose their edge for frontier training within 3 years, but are being depreciated over 5-6 years. He estimates this resulted in profits overstated by $176 billion.

The evidence is thin. Amazon actually shortened its useful life estimates in 2025 and Meta extended theirs. Nvidia argues that chips are still running at full utilization after six years. GAAP accounting rules give companies a great deal of discretion in how they report this. So it’s unlikely to be a sign of fraud.

Negative free cash flow has been central to the debate. Alphabet’s free cash flow turned negative for the first time since its 2004 IPO. Bank of America forecasts hyperscalers’ cash flow dropping from +$180 billion in 2025 to -$64 billion in 2026, then to -$144 billion in 2027 and -$186 billion in 2028.

But negative free cash flow is not inherently a bad thing. It means these companies are investing more than operations are currently bringing in. Sophisticated management teams don’t shovel hundreds of billions into infrastructure without having a strong reason to believe it will pay off. There are only a few uses for free cash flow (invest in growth, return to shareholders, or pay down debt), and investing aggressively in growth is one of the most bullish signals there is. Whether it results in the ROI they are hoping for remains an open question.

The hyperscaler build out is circular. Nvidia spends $100 billion with OpenAI, OpenAI spends $300 billion on a cloud contract with Oracle, and Oracle buys chips from Nvidia. Layering debt on top of this cycle makes it more fragile. Oracle carries negative $24 billion in free cash flow against $219 billion in liabilities.

Overvaluation most likely exists in the private markets. AI companies with no revenue are already being valued in the billions.

Disillusionment settles in

Hyperscaler capex has climbed to almost 100% today, from its normalized average of 40%. Capital allocators are betting that the demand eventually shows up. Here’s why it might not.

• People overestimate AI’s impact on their tasks by 40 percentage points.

• 76% of companies believe they are ahead of their competitors on AI, but only 10% are seeing real ROI.

• 95% of generative AI pilots show no measurable financial impact.

• A 2026 survey of 2,850 business leaders found that those who described their AI usage as “aggressive” declined from 60% to 42% y/y.

• Nearly 70% of executives said they would cut their AI budgets if ROI targets were not met.

• 74% of companies that deployed AI agents in customer communications have rolled them back.

AI applications that have directly measurable impact are still likely to get funded. For example, coding tools are driving Anthropic’s growth.

Even if AI succeeds

There’s a separate reason a market correction could happen. AI’s uncertainty is a risk increasingly shared by the whole economy. It is not contained by just a few companies. A greater level of risk implies investors will demand a higher premium for holding it. That alone could pull valuations down. In this scenario, a market correction may happen even if everything goes well.

What to watch

Hyperscaler capex continues to climb while the customer demand is pulling back. Customers are capping budgets and cutting projects that don’t deliver measurable financial impact. If that gap doesn’t close, the first casualties will be companies with huge valuations but no revenue and companies that are over-leveraged.

On the other side, some profitable SaaS companies look underpriced on AI fears that haven’t, and perhaps won’t, materialize. More on that later – subscribe so you don’t miss it.

The tool itself was simple to build. AI excels at running complicated analyses with speed.

The input assumptions for the analysis were much harder to pin down. They are critical for getting an output that is meaningful and usable. Garbage in, garbage out. Claude confidently stated wrong assumptions and changed its answer every time I questioned it.

Still, I found the analysis useful. It helped me think through how I manage risk in my portfolio. It reminded me why it’s important to focus on risk-adjusted returns rather than highest raw returns.

How to build it

I started by uploading a spreadsheet that contained my current portfolio holdings.

I prompted Claude to categorize the holdings into specific asset classes. Differentiate between treasury bonds, corporate bond, and securitized products, rather than combining all types of fixed income instruments.

I prompted Claude to create a tool with adjustable toggles. Each toggle represents an asset class’s percentage of the total. Users can adjust these percentages in either direction using manual controls. The tool applies assumptions for the asset classes’ expected returns and volatility.

The tool will re-run the analysis based on the adjusted allocations. It will then show the expected returns and volatility of the total portfolio. The tool uses a Monte Carlo-like simulation. It projects thousands of possible outcomes. The market is unpredictable. A Monte Carlo analysis aims to model the range of potential outcomes.

It also has a field where users can enter how much new capital they will add to the portfolio annually.

The output included a graph. It showed how the portfolio would grow over 10, 20, and 30 years. All modeled scenarios formed the basis for this. It showed a side-by-side comparison of the expected returns. One was for my current portfolio, and the other was for the new portfolio with my changes.

The shaded blue range illustrates the 90th to the 10th percentile outcomes. The dark blue line indicates which outcome is the most likely. Compare this the gray line, which indicates the trajectory of the current portfolio without modeled changes.

The output showed the portfolio’s geometric annual growth rate and volatility. From those, it projected what the portfolio would grow to in a bad market, what it could grow to in a good market, and what it’s most likely growth would be. Each shown for both the current and adjusted portfolio.

The assumptions are critical

To run this analysis, Claude had to estimate the expected returns and volatility for each asset class. This is the part that is a bit subjective. Every time I asked it to verify these assumptions, it changed them. At one point, it sourced estimates from Charles Schwab. I asked it to perform its own analysis and produce reasonable assumptions. It kept changing its mind. Assumptions about the future are, by definition, unknowable. Yet, the output that comes out of this analysis is only as solid as the inputs that go into it.

In the last decade, U.S. stocks did much better. In contrast, international and emerging markets lagged behind. The U.S. is a more mature market, and it is generally expected to have a slower growth rate. The reason the last decade was so unusual is that a small group of Tech/AI stocks carried the U.S. market. The “Magnificent Seven” accounted for roughly 42% of the S&P 500’s total return in 2025 alone. Most companies thriving from the AI boom since ChatGPT’s 2022 launch are based in the U.S.

Many forecasters expect slower growth in U.S. markets over the next decade. This is because valuations are much higher than usual. (Of course, back in 2017, Vanguard predicted a decade of low growth, and we know how wrong that was.) The gap between U.S. valuations and those in international or emerging markets is large. It’s the largest it has ever been. Which geography is best positioned to gain from the next decade of AI-driven growth? There is no clear answer. This is another key assumption in the model that is almost impossible to determine.

Here are sources to inform model assumptions:

• Vanguard Capital Market Model forecasts (published June 2026)

• JPMorgan Long-Term Capital Market Assumptions (published Sept 2025)

• GMO 7-Year Asset Class Forecast (published June 2026)

• BlackRock Capital Market Assumptions

Geometric returns and volatility

Once I had the tool set up, I could do the more interesting analysis. I asked Claude to reverse-engineer the construction of the highest growth portfolio. Claude ran the calculations and showed me the optimal way to achieve high growth. One change it recommended for my portfolio was reducing volatility.

Studies say that women tend to make better investors than men. Women are generally more risk-averse. Over time, this tends to result in higher returns. Men often look for higher volatility in their investment strategies. They trade more often and tend to hold on to losses longer. The original study found that men trade 45% more than women. Women increase their returns by trading infrequently and cutting losses sooner.

The Portfolio Allocation Lab modeled this. Increasing volatility levels raise the best-case scenario while also lowering the worst-case scenario. The most likely scenario often becomes worse. The potential deeper losses pull down the median outcome.

This is a good spot to explain a key financial idea. Let’s look at the difference between arithmetic returns and geometric returns.

Arithmetic returns are simple averages, while geometric returns account for compounding over time. The geometric return is the more accurate calculation.

Imagine playing a game of coin flipping. The probability each time is 50%. Start with $100. In year 1, you win the coin flip and gain 50%. You now have $150. In year 2, you lose the coin flip and lose 50%. You may think “+50% and -50% cancel out, so I’m back to $100.” This is incorrect. Losing 50% of $150 sets you back to $75. Even though your two yearly returns “average” to zero, you lost 25% of your initial $100.

Two portfolios can have the exact same average annual return. However, they can end up in very different places due to how bumpy the ride was. The volatility and fluctuations in value make a big difference in the final outcome.

• Steady: +5%, then +5%. $100 → $105 → $110.25. Total gain: +10.25%

• Bumpy: +30%, then −20%. Average of +30 and −20 is still +5%. But: $100 → $130 → $104. Total gain: +4%

Concentration vs. diversification

The Portfolio Allocation Lab hinges on diversification. My personal portfolio has several single-stock bets (because I like picking stocks, sue me).

Many investors will attribute large investment fortunes to concentration. In other words, going all in on one bet. Concentration widens the range of outcomes. A few end up with massive fortunes, and they are outliers. Most end up much worse off, and their stories are rarely told.

In 2018, Hendrik Bessembinder studied all U.S. stocks back to 1926. Only 4% of stocks drove the market’s outperformance of Treasury bills. The majority (51.6%) of stocks lost money. Concentration bets you can pick that 4%.

A concentrated bet only reliably pays off when you know something the market doesn’t. Without a real edge, it’s just uncompensated volatility.

Investor’s edge

When I pushed Claude to find higher growth, it directed me to leverage.

Cheap leverage is hard for individuals to find. Schwab margin: ~10%. Interactive Brokers: ~5%. Both risk margin calls, and most other credit lines bar you from investing the proceeds.

One option it directed me to is ETFs with leverage built in via Treasury futures. $1 invested buys $1 of global equities plus $1 of Treasury exposure, at cheaper-than-retail financing. Launched only in 2023, so data is thin.

Warren Buffett is basically a case study in this. AQR's "Buffett's Alpha" paper found Berkshire's edge wasn't really concentration. It was ~1.6x leverage, financed through insurance float.

Try it yourself

If you would like to complete this portfolio exercise on your own, here is the prompt I used to build the tool and run the analysis.

“Review the attached positions in the [Portfolio Positions] document. Create a tool that models future returns based on allocations across asset classes. Organize the attached holdings into each category. Be precise on the categories (not simply “fixed income”, but differentiate between treasuries, corporate bonds, securitized). For each asset class, define return and volatility assumptions. Build a correlation matrix between the asset classes. The simulation should account for how they move together. Build a toggle for each asset class where I can adjust the allocations up or down. Run a monte carlo-like simulation. Create a graph that displays the 10-year, 20-year, and 30-year views. Include the 10th and 90th percentile bands along with the median (geometric) outcome. I want to be able to adjust the individual asset class toggles and see how it would impact future portfolio growth.”

Don’t get me wrong. There is a lot we can learn from him. He has shared much wisdom that still holds true today.

He preached stoicism: a trait many investors would enjoy adopting.

• “If you can’t control your emotions, you can’t control your money.”

• “Be fearful when others are greedy and greedy when others are fearful.”

• “The most important quality for an investor is temperament, not intellect.”

But study the methods he used to achieve his success, and you may find the opportunity has evaporated.

His track record

Buffett graduated from University of Nebraska in 1950 with a B.S. in Business Administration. He then earned an M.S. in Economics in 1951 from Columbia Business School. He went to work at father’s firm, Buffett-Falk & Co. as a securities salesman. After a few years, he returned to New York, taking a position at Graham-Newman Corporation as a securities analyst.

In 1956, Buffett returned to Omaha and started his own investment firm. He raised $105,100 from family and friends, plus $100 of his own money. Additional single-family partnerships brought the total assets under management to $303,726 ($3,674,997 in today’s dollars).

During those early partnership years, he achieved returns of ~29% annually. After taking control of Berkshire Hathaway in 1965, Berkshire compounded around ~27% from 1965 through 1999. For the last 20 years, he’s roughly tracked with the S&P 500.

The edge behind his returns

Investment research was a different animal pre-2000. An analyst would simply go meet with management, ask how business has been lately, and then use that information to inform their trades. They called this method “scuttlebutt.”

Passed in 2000, Regulation FD banned companies from feeding information to some analysts but not to others. Today, the market has rules around public dissemination. When a CEO shares information about how business has been lately, they must say it publicly so all investors hear it at the same time.

This part of Buffett’s strategy – the access edge – cannot be replicated in today’s stock market.

He also traded on an obscurity edge. Buffett has said if he were working with $1 million today, he could “guarantee” a ~50% annual return, because tiny, illiquid, uncovered stocks are where real mispricings persist. That’s an advantage available to any individual investor willing to put in the research. The whole small-cap market is open to you in a way it isn’t to large funds.

He was known for hand-collecting data on tiny stocks that no one bothered looking at. This could still be replicated today. You can run a screen for micro-cap stocks too small for institutions to own and too small for analysts to cover. However, less asymmetry exists today than in 1950. Online stock screeners make this easily visible to everyone, thus eliminating most obvious arbitrage opportunities.

The other edge Buffett possessed is a behavioral one. This one is fully replicable.

What his letters actually teach

If you read the letters to shareholders written by Buffett, you would walk away with these core principles.

1. Invest in excellent businesses at good prices. Don’t invest in mediocre businesses at great prices. Quality is essential.

2. Temperament beats intellect. Control your emotions, or else your emotions will control your money, and the market will punish that.

3. Compounding requires doing nothing. Place trades infrequently and be patient for decades. This is where the magic happens.

He is a big fan of durable competitive moats, high returns on capital, honest and able management, understandable businesses, real cash generation, buying with a margin of safety, extreme concentration, and near-permanent holding periods.

Should you copy his portfolio?

If you sat down with Warren Buffett and asked him for investing advice, he wouldn’t point you to his portfolio as inspiration. He would not hand you a strategy for picking or analyzing stocks.

He would tell you to invest in a low-cost index fund.

I agree that, for most people, low-cost index funds are the right strategy. Picking stocks is a full-time job and most underestimate how time consuming it is to do it properly.

I’d also caution that there is no such thing as a perfect investment. There are only those that fit your risk tolerance, time horizon, and goals better or worse. Buffett is 95 years old and has a net worth of $147 billion. His strategy is naturally going to look different than, say, a 30-year-old investing $100,000. He has a much different objective – capital preservation and legacy, not growth. That’s why his portfolio holds stable, large cap companies with dividend payouts. That same portfolio would likely make far less sense for someone younger, whose main objective is growth.

Buffett has long had a reputation for avoiding investing in technology companies. His philosophy is to stay within your circle of competence and only invest in businesses you understand. He was slow to invest in Apple, which ended up being his best investment of all time once he eventually bit the bullet. He missed out on companies like Nvidia, which grew 87x over the last decade, and Tesla, which grew 44x. Avoid tech investing at your own peril.

Learn from his philosophy, but don’t copy his strategy.

Other investing role models

There’s more wisdom out there than one shareholder letter collection. Here’s a list to get you started.

1. Lee Freeman-Shor - The Art of Execution

This book analyzes what elite fund managers’ actual trades revealed vs. what their stated philosophies were. It found that position management mattered more than idea quality. It teaches concrete rules for adding, holding, and exiting a position.

2. Chris Mayer - 100 Baggers

A systematic study of stocks that returned 100x outcomes. It teaches the statistical profile of companies early enough to compound 100x (small, high ROIC, reinvestment runway, owner-operators), and why selling winners too early is an expensive habit.

3. Carlota Perez - Technological Revolutions and Financial Capital

The theory behind megatrend investing. It outlines the phases of a big technological advancement, from irruption to frenzy to crash, and eventually to maturity. It’s useful for placing any technology, AI included, on that cycle. Her papers and talks are available here.

4. Howard Marks - Memos + Mastering the Market Cycle

He teaches you to think in probabilities and find where you are in the market cycle. Memos are available for free at oaktreecapital.com.

5. Nick Sleep - The Nomad Partnership Letters

His fund invested in 2-3 retailer stocks, when everyone else believed they were overvalued. He held through the noise and stood firm in his conviction. Once he ran out of new ideas, he closed the fund. A full collection of his letters is available here.

6. Stanley Druckenmiller - Interviews & Speeches

His investment style is very flexible. He will invest in currencies, bonds, or equities depending on what the macro backdrop rewards. Here’s a more recent long-form interview he gave.

7. Li Lu - Columbia Lectures + Essays

As a modern-day value investor, his hedge fund has reportedly returned a 30% compound annual return since 1998. His Columbia University lectures are available on youtube.

8. Annie Duke - Thinking in Bets + Quit

As a former professional poker player, she teaches you to separate good decisions from good outcomes. You can make the right call and still lose, or the wrong call and still win. Also makes the case that quitting on time is a skill, not a failure.

9. Nassim Taleb - Fooled by Randomness

Rare, high-impact events dominate outcomes far more than most people’s models assume.

10. You - Yes, you.

One of the best ways to learn and improve your own style as an investor is to keep a written journal on your trades. Document your thesis. Analyze your errors and your wins. Write an annual investor letter to yourself.

Buffett used to ask business students to perform a mental exercise. You can buy 10% of the lifetime earnings of one of your classmates. Who would you choose? What traits does that person possess that makes you choose them? Perhaps characteristics like integrity, generosity, or good judgement. Then he would point out that these are traits you can nurture in yourself. You owe it to yourself to become the person worth betting on.

So when it came time to choose a major in college, it was no surprise that I chose “playing with money” as my major. In other words, I chose finance.

Our campus investment club held a weekly career series. Finance professionals from different local companies came to share their advice. They shared what a typical day in their job looked like. Then, they pitched their company as a great place to work.

Investment bankers came to speak at our college. They told us that investment banking is the most lucrative career we could choose. They told us that the only 22-year-olds who earn more are famous celebrities and pro athletes. Naturally, I was sold.

In college, we learned how to build a discounted cash flow model (“DCF”). We learned that this was the best method for determining the value of a company. Professors endorsed it as the most scientific method.

Fast forward to my first day working on Wall Street. We were kicking off the roadshow for an IPO. A pitch deck was on my desk, along with instructions to memorize it inside out. Imagine my horror as I flipped through the pitch deck and found not a discounted cash flow model in sight.

I would listen in on calls between the trading desk and hedge fund managers. The only topic up for debate is what multiple a company deserves. Multiples are much simpler than the complex financial models I had trained to build. I pondered how divorced the price of these stocks may be from their intrinsic value.

Of course, the biggest problem with DCFs is that the answer they spit out is only as good as the inputs that go in. Garbage in, garbage out. The same is true of large language models like ChatGPT.

When I was in college, I came across research that almost ended my career before it even started. It was about alpha. Alpha is the return you make above and beyond what the market gives you. If the market returns 5% and your portfolio returns 7%, then your alpha is 2%. (I’ve always thought consistently underperforming the market is its own unique skill. Perhaps that should get its own Greek letter too. I like to call it “omega.”)

The research found that when an investment manager beats the market one year, there’s a 50% chance he’ll beat the market next year. He has no better probability than a coin flip. Meaning, there is no such thing as skill in investing. There’s only luck.

Over any long stretch, almost no one beats the market. (Not even Warren Buffet -- more on that next week).

In 2021, 20.1% of large-cap funds in the top quartile stayed there in 2022. By 2023, that number dropped to 0.0%. In more than 15 years, about 90% of professional managers did worse than their index.

So Alpha is like Bigfoot. Have there been any confirmed sightings? Everyone has a cousin who saw him, but nobody has a clear photo.

So if it’s mostly luck, and not even the top professionals can beat a plain index fund, then why not just buy the S&P 500, close your laptop, and go outside?

The closer I got to Wall Street, the worse the picture became. I saw how wide the gap is between regular people who invest and professional investors. Most regular people who invest underestimate:

1. The sheer time it takes to cover one stock properly. It’s a full-time job for a team of analysts, and then some. You can’t follow just one company; you also need to track its public and private competitors

2. The depth of expertise the analysts have. i.e., most biotech stock analysts I know have PhDs in specialties of science I didn’t even know existed

3. The information gap, which is huge. Professional investors sit down face-to-face with the CEOs. Regulations prevent those CEOs from sharing anything they haven’t already made public. But there are no regulations around body language. Communication gets conveyed. If those meetings were worthless, they wouldn’t keep happening.

4. Markets are efficient. Pretty dang efficient, from where I was sitting. By the time you read an article about a company, whatever it says is already reflected in the stock price.

In 2022, I bought every quantum computing company that was publicly traded. I had a long-term thesis on the technology and planned to hold them for 20 years. (Patience is my investing strong suit.)

My thesis was wrong. It didn’t take 20 years; it took three. The quantum stocks I held had run up 50x by 2025. The valuations made me nauseous.

The market was wrong both times. The stocks were priced for dead when I bought them. And when I sold, they had priced in a future that hadn’t remotely shown up yet. Markets may be efficient, but they aren’t always rational. They will go from ignoring a sector to obsessing over it, and then back again. This is also how I made 34x I made on a dying crypto miner that everyone else had written off.

I’ve beaten the market plenty of times through a combination of skill and chance. Some of it is that I’ve picked good hunting grounds.

A good hunting ground is the patch of the market that fits the game you’re actually playing. I have a long time frame and can wait for years without seeing action. So, I seek out emerging technologies that have the potential to make a significant impact in the future. I wasn’t buying Coca-Cola. Those are good grounds, and suitable for somebody else.

Here’s why I keep saying this: don’t get investing advice from someone who doesn’t know your situation. There are no perfect investments, only those that fit your risk tolerance, time horizon, and goals better or worse. The best hunting ground for me might be the worst for you. No one can know what you need unless they understand your goals and limits.

Here’s why I keep at it despite everything above. The stock market is one of the few places you can access limited downside and unlimited upside at the same time. You can only lose what you put in. The wealthiest people already understand this. For a middle-class family, housing and pension accounts make up about 80% of their net worth. Stocks make up only 4%. For the richest 1%, their primary residence is a mere 7.6% of assets. Pouring your whole net worth into the house you live in is a middle-class move. The rich hold businesses and stocks, ownership of things that compound without them. You can argue chicken or egg about which came first. The association is too strong to wave away.

So should you try to pick stocks? For most people, no. Buy the index, automate it, and don’t spend your weekends on this.

But if you’re going to do it anyway, and some of us are, here’s what you should know. The edge is real, and it’s not for sale. An AI chatbot hands you the consensus. It doesn’t have information the market doesn’t already know or suspect. To find significant mispricings, you need to know something the market doesn’t or believe something it hasn’t yet accepted. More on how I find these spots every week. Subscribe to follow along.

This business model is under attack from three directions (regulation, economics, and AI). It may not survive in its current form.

Defining the players

The sell-side includes banks and brokerages like Goldman, Morgan Stanley, and Jefferies. They provide research, trading, and deal services to investors. A sell-side equity research analyst covers about 15 to 25 stocks in one sector. They publish opinions on these stocks. Their clients are the buy-side money managers.

The buy-side includes institutions that manage money and buy securities, like stocks. These are hedge funds, mutual funds, and pensions. Buy-side analysts and portfolio managers decide what to buy with their funds’ money.

The revenue streams

The broker vote

Pricing is bespoke and negotiated per client. The institutional investors set aside a discretionary budget for research each year. At the end of the year, they do a broker vote. Their investment teams vote on which sell-side research analysts provided the most value. Then they dole out dollars accordingly.

Corporate access is the single biggest monetization lever. It often drives the outcome of the broker vote more than any written research reports. Corporate access means meeting with a company’s management team. Sell-side analysts help make these connections. It includes conferences, non-deal roadshows, and 1:1s with CEOs. Consider the incentive this creates for sell-side analysts. Their compensation rides on whether the CEOs of the companies they cover like them. Not very unbiased, I’d say.

Investment banking transactions (indirect)

In 2002 and 2003, lawmakers passed regulations after the dot-com scandals. These rules separated investment bank transactions from the pay of research analysts. The investment banking team can still influence which analysts are most valuable to the firm. They base this on the deal pipeline. Bankers want to pitch private companies on having the best analyst in their industry.

Trading commissions

Trading commissions have largely collapsed as a profit stream. Most trades are now done electronically. They cost much less than the fees high-touch trading desks used to charge. The rise of passive capital has also resulted in less active trading and less research consumed. The trading commission pool still exists, but at a fraction of its former size.

Regulation tried to unbundle the business model

Regulators ran the experiment of forcing this business model to change. In 2018, Europe’s MiFID II rules made investors pay for research separately from trading. This resulted in research budgets shrinking, analyst headcount falling, and coverage of smaller companies vanished. It went badly enough that regulators began walking back the regulations. Now, European firms can bundle research and trading payments once again. The lesson learned is that when clients are forced to put an explicit price on sell-side research, they decide it isn’t worth much.

What the buy side wants to pay for

Buy-side investment teams dismiss sell-side analysts’ price targets and ratings (buy/hold/sell). Any half-decent buy-side firm runs its own analysis and has its own thesis. They are, however, willing to pay for sell-side research that offers proprietary insights.

This research might use satellite images of a retailer’s parking lot. It could also include credit card transaction data and interviews with customers and suppliers. As a sell-side research analyst, I called STD clinics to assess volume trends for various competitors, had blood drawn at three different labs in just one day, and read the serial number off the back of a new sequencer in a genomics lab.

AI is commoditizing research

The proprietary boots-on-the-ground research that once gave an edge is now being leveled out. Buy-side firms are increasing their spend on alternative data and using genAI to analyze it efficiently. Citadel and BlackRock have added genAI tools to their platforms. Large funds are creating their own systems. They are using private LLMs and training them on licensed data sets.

The buy-side is doing its own research in a more sophisticated way than before, using AI agents at little marginal cost. The market has commoditized the last differentiated product of sell-side research. This leaves only corporate access as their final moat. Equity research becomes little more than a relationship broker.

What this means for the individual investors

An individual investor can use AI to replace a junior investment analyst. AI can read and summarize a 10-K quickly and for minimal cost. This provides no edge to institutional investors who have access to the same AI tools. When everyone has AI, AI is not an edge but table stakes.

Proprietary research and alternative data often influence stock prices quickly. This happens before individual investors have a chance to act. AI doesn’t equalize access to information. Institutions still have better access to data.

But individual investors have edges that institutions can’t replicate, and AI amplifies them. An individual investor can own small-cap stocks too illiquid for a billion-dollar fund to touch. They can hold through a drawdown with no career risk and no quarterly benchmark to answer to. And now, with AI, one person can do the processing work of a junior analyst on companies too small for Wall Street to cover at all. AI plus small caps plus patience is a real strategy.

This period has made a straightforward investment strategy popular: “VOO and chill.” It’s easy enough to fit on a bumper sticker. You buy and hold a single index fund that tracks the market. If you followed this strategy, you would have averaged a 15% gain per year. Compare this to the long-term market average of 10%. It may sound like a minor difference, but the compounding over time is what makes it add up. It’s the difference between $1,000 growing to roughly $2,600 vs. $1,000 growing to roughly $4,200.

Growth Driven by a Select Few

The abnormal returns came from a small group of American tech companies. These are Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, and Tesla. In 2015, these seven made up about 12% of the S&P 500, with a combined market value of around $2.2 trillion. Today, their combined value exceeds $22 trillion, which is roughly a tenfold increase. They account for about a third of the entire index.

The chart needs a log scale to fit on one axis. Nvidia grew around 87 times, and Tesla about 44 times. The slower members still grew five to eight times.

Companies rise and fall. That’s the natural life cycle of business. But consider how rare these seven businesses are. The scale of these businesses is without precedent. The companies at the top of today’s market report record annual profits. These profits are the largest ever seen by any public company.

The other 493 companies of the S&P 500 grew at a historically normal rate during this period, roughly 9% per year. The whole abnormality revolves around seven names. The wealth created by the VOO and chill crowd depends on them.

Prices Increased

The other factor that drove abnormal returns was multiple expansion. Investors now pay more for each dollar of earnings than before.

The Shiller CAPE index compares market prices to inflation-adjusted earnings. Since 1881, its long-run average has been 17x. Today, the reading is 42x. Investors are now paying about 2.4 times the usual amount for each dollar of earnings. In 145 years of data, only December 1999 showed a higher rate. That was months before the index dropped by nearly half.

Part of what pushed prices up is the supply of money flowing into the stock market itself. Between 2016 and 2025, passive funds in the U.S. took in $6.4 trillion of new money while active funds bled out $2.4 trillion. In total, about $4 trillion of new capital flowed into the market over the decade. This reflects both a rotation and an expansion. This shift increased index funds’ share of equity fund assets from 36% to 57%. This affects prices because passive capital doesn’t care about value. It doesn’t ask if Nvidia is cheap; it just buys what the index includes, regardless of what the price is.

Some of the price expansion is fair. In the 1930s, the stock market was railroads, steel mills, and other capital-intensive industrials. The stock market today is higher quality. It consists of scalable software platforms with high profit margins. A dollar of software earnings deserves a higher price than a dollar of railroad earnings.

How much more could multiples realistically expand from here? By any historical measure, almost none. At 42x, the CAPE has been exceeded only once, right before a massive crash. Goldman Sachs predicts the S&P 500’s forward multiple will drop from 23x to 21x by 2035. They expect the U.S. stock market to return 6.5% a year. This is much lower than the 15% annual returns seen over the last decade.

Keep in mind, the CAPE has sat above its long-run mean almost continuously since 1991. Anyone who acted on the belief it’s overpriced would have missed the best thirty-year run in market history.

International Stocks Sat This One Out

Over the last decade, international stocks significantly underdelivered relative to the gain of U.S. stocks. The valuation gap that opened up along the way is now the widest on record. The S&P 500 trades around 21.5x forward earnings, non-U.S. stocks trade near 15.5x, and emerging markets sit at 13x.

Cross-country CAPE comparisons are not always reliable. Different accounting, different sector mixes, different governance, and cheap markets are frequently cheap for excellent reasons.

The explanation for this gap is primarily sector composition. International markets did not enjoy the AI boom the way the U.S. did. A J.P. Morgan analysis found that since ChatGPT launched in November 2022, AI stocks accounted for 75% of the S&P 500’s returns. They also contributed to 80% of the earnings growth. Taiwan’s chip fabs and Korea’s memory makers were unique cases positioned to benefit from this boom.

Claude Fable 5 is the new top of Anthropic’s lineup. It ranked highest on Hebbia’s financial-reasoning benchmark. IMC, a trading firm, noted it outperformed all their internal trading tests. It went live in June, got pulled offline three days later by U.S. export-control rules, and came back online last week.

Fable checks each request for risks, like cybersecurity and biology. If something raises a flag, it passes the session to Opus, a less powerful model. Anthropic says that happens in under 5% of sessions. Mine ran closer to half, sometimes on a prompt that had worked minutes earlier in another window. Anthropic confirmed that a portfolio review is allowed for Fable. I still don’t know what triggered this.

To start, I loaded a CSV of my stock and ETF holdings and gave Fable my age, filing status, net worth, income, and goals.

Generic advice isn’t advice

Be cautious with investment advice from anyone who doesn’t understand your financial situation. Their guidance may not be right for you. There’s no perfect investment. There are only those that fit your risk tolerance, time horizon, and goals better or worse. Advice from someone who doesn’t know your situation is close to worthless.

So before Fable analyzed anything, I had it quiz me to determine my risk tolerance. What people say they’ll do in a crash and what they actually did in the last one are usually different answers.

Fable’s quiz focused on behavior more than self-report. It asked what I did the last time the market dropped 40%. It also wanted to know why I’m still holding the bond funds. It asked which would hurt more: a position going to zero or missing a stock on my watchlist that tripled. Then it checked my answers against my cost basis to see whether I’d held through the lows I claimed I’d held through. It graded my trading history instead of my story about myself.

Checking the holdings against the goal

Once it had a read on my risk tolerance, Fable went through the holdings against my stated goals. It marked my bond and commodity allocation as too high for my goals. It also urged me to sell the positions that were too small to matter.

Plotting the portfolio by risk and return

Then I asked it to plot every position on a grid, risk on one axis, return potential on the other. I ran the same prompt twice in separate windows to see whether the map stayed the same.

I got a different map each time, but much of the map held. My large-cap software names sat in the low-risk, modest-return corner both times. The profitless quantum-computing and gene-editing names came back high-risk on both passes. The green cluster Fable read as limited downside with real upside didn’t move. Where the two maps split was a dozen or so names sitting right on a dividing line. CRBU and GUTS dropped out of the high-reward quadrant into the limited-upside one on the second run. FIG went the other way. AXON and PLTR crossed the risk line. MNSO, which sits near both lines at once, showed up in the sweet spot on one map and the danger zone on the other.

These models are non-deterministic. This means that the same prompt generates scores from scratch each time, resulting in slight differences. If Fable scores one ticker 5.6 on return one run and 5.4 the next, it falls on opposite sides of a line drawn at 5.5. It lands in a different quadrant even though the underlying read barely changes. That wobble is a list of the calls that are close, and those are the only ones worth your time. Fable can place the obvious names on its own. The ones that jump between runs are where the market hasn’t settled either, and where your own work has to happen.

Everything on these maps is a real position in my brokerage account. I’m showing you my book to make a point about the exercise, not handing you a buy list. What lands in my sweet spot could be wrong for yours.

High risk does not equal high return

Alarmingly, the analysis showed that 40% of my capital is in the high risk, low return area. This is an error that’s easy to make: taking on risk with the expectation of higher returns. Many people mistakenly assume risk and return are on the same dial.

Risk and return are on separate axes. High risk widens the range of possible outcomes. However, it does not say whether the top of that range is any good. A pre-revenue biotech burning cash against three competitors with a dilution problem on top is about as risky as a position gets. Its realistic upside can still be mediocre once you count the ways value leaks out before you ever see a dollar of it.

If you want to try this

Fable scores these names off what’s already been written about them, so the map it draws is the market’s current read on my portfolio and little else. I made that case in AI Will Make You a Better Investor, Not a Great One: the model gives back the consensus, and a consensus that’s already in the price isn’t an edge. A grid like this is good for seeing how my money lines up against what everyone believes and no help in finding the spot where everyone’s wrong. That part’s still mine, which is why I stay suspicious when the model likes everything I own.

This persists across all AI models. In 2023, Anthropic released a study titled, “Towards Understanding Sycophancy in Language Models.” It tested five top assistants from different labs. The study found that all of them displayed sycophancy across various tasks. In April 2025, OpenAI rolled back a version of GPT-4o that had become so eager to please it was praising obviously bad decisions and going along with claims it should have questioned. Yikes.

Why it happens, and why it isn’t going away

An LLM is a next-word prediction engine. It read an enormous amount of text and learned to continue a passage in the most plausible way. By itself, this creates fluent but unfocused results. So, labs add a second stage. People compare the model’s answers and rate the ones they prefer. Then, the model gets tuned toward whatever scores well.

That second stage is where the problem lives. People prefer agreeable, confident, flattering answers. They rate those higher than blunt or contradictory responses. So, training often rewards agreement. Anthropic traced the behavior straight to that preference data. They found the rating systems often favored a convincing wrong answer over a correct one. When the model agrees with you, it is simply behaving according to its training.

What it means for investing

Sycophancy may be a minor annoyance for most things you’d ask a chatbot, but it’s a much bigger problem when there’s money on the line.

The model will say a stock is cheap. If you return next week and say it’s expensive, it will agree again. It won’t remember the earlier view. It holds no position, so it has nothing to defend. The way you make money in individual stocks is by knowing something the market hasn’t priced in yet. That needs two things: a correct view and the belief to stick with it. This is something an overly agreeable AI model can’t help you with.

I made 34x on Applied Digital after sitting with a thesis for a year and a half. The consensus rejected it the entire time. I wrote about this in I Made 3,388% on a Single Stock. A good AI model would never have picked that stock because the market didn’t see its potential. AI only reflects what the market has already decided.

How to actually use it

None of this makes the tool useless. It’s a tool that becomes highly useful once you learn how to use it.

The most valuable thing a consensus machine can do for you is show you the consensus. I wrote about why these tools return the market’s existing view in AI for Investment Research, and that’s the feature here, not the bug. Ask it for an analysis of a company, and it will deliver a clean map of what’s already priced in. Your job isn’t to accept that map; it’s to find the wrong assumption that everyone believes is true.

Three ways to get real value out of it:

Use it to surface what’s priced in. Lay out the bull and bear cases that the market holds. Then, view these as beliefs to challenge, not to confirm.

Make it argue against you. Tell it you hold the opposite of your real position and ask for its strongest work. A bear case it can’t articulate is a signal that the downside may be thinner than it looks.

Point it at checkable work. Upload the filings. Use them for retrieval and synthesis. The document should guide the answer without any embellishments.

Three things to watch:

Agreement is not a second opinion. If the model supports your thesis, it means you crafted a strong prompt and nothing more.

It can’t do the arithmetic. It will set up a DCF correctly and confidently drop a digit in the middle. So, recompute anything that influences a decision.

It doesn’t know what matters. It weighs every line of a filing equally and may present old facts as new. This means it’s up to you to judge what’s important.

I've been using AI for investment research for almost a year now. I've written about the obvious errors before, the bad math and the stale numbers it reports with a straight face, in Three Things AI Will Confidently Get Wrong About Your Investments. AI Will Make You a Better Investor, Not a Great One was about a subtler problem: everyone landing on the same answer because they're all asking the same models the same questions. An agreeable read on a stock is worthless when it's the consensus that's already in the price. If you'd like to learn more about my investment process and how I use AI to find what the market's already missed, subscribe.

I spent years in equity research, and for most people the answer is no. It can be done. The hours it takes rarely clear the bar once you price them honestly.

Count the hours

Tracking companies is like having a second job. You read filings, build models, and listen to earnings calls. You also update your thesis whenever the facts change. Six hours a week is three hundred hours a year, and that’s conservative.

Now look at what those hours buy. Say you’ve carved out $100,000 for individual names on top of a diversified base. If you have a great year and double it, you’ve made $100,000 before tax. Then divide it by three hundred hours, and then by all the years you didn’t double it, and a different picture emerges.

Most people researching stocks are valuing their time at zero and never saying so out loud. The same hours aimed at your income or something that compounds without you would build wealth in a way a stock slice structurally can’t (the asymmetric bets framework).

There’s also a cost that shows up in April. Sell a winner you’ve held under a year and the gain is taxed as ordinary income, a third or more of it for a high earner. An index fund you don’t touch delays taxes for years. When you sell, you pay the lower long-term rate.

You don’t have the information the pros have

The hours would be worth it if the odds were good, but they’re stacked against you, and the reason is information.

When I was on the sell side, the returns our institutional clients paid for didn’t come from public filings. They came from proprietary work: channel checks, expert calls, data nobody else had. I crawled under a DNA sequencer in a laboratory to read a serial number. This helped us estimate a company’s quarterly shipments (I Pre-Wrote My Research Reports Before the Earnings Call Even Happened). That’s the level of effort on the other side of your trades.

By the time you’ve read an article about a company, the people who move the stock have already priced in what it says. Despite all that, professionals often fall short. Over fifteen years, fewer than one in ten beat the index they’re measured against. You’re competing with them for an edge they can’t reliably find themselves.

I’m not going to tell you not to pick stocks

None of this makes individual stocks off-limits. I pick them, and I’ve written about a position that returned more than 30x what I put in. One big win is just an outcome, not a track record. Thinking a good result proves you can repeat it is the quickest way to lose money in individual stocks.

So the honest question isn’t whether you can pick stocks. It’s why you’re doing it. A few reasons hold up. Enjoying the analysis is one, as long as you call it a hobby and pay for it like one. Learning the craft is another, though the early years are tuition and you should treat them that way. The last is having a real edge, which is rarer than nearly everyone who claims it believes, and which no run of winning trades will ever prove to you. The reasons people fail are often the same: they think picking stocks will make them rich or they see someone else’s big gains and want the same. The costs above are the price of those reasons, and they fall hardest on the people least able to tell which reason is theirs.

Be honest about which reason is yours, keep the slice small enough that being wrong costs you nothing important, and put the rest of your attention where it moves the number.

The gains weren’t even. The weakest writers improved the most, while the strongest didn’t write any better with AI at all. They simply finished faster. The tool pulled the bottom toward the middle and left the top where it was.

I’ve used AI to research and manage my own portfolio for the past year. It’s fast at the busywork and useless at the only thing that pays: catching what the market got wrong. Now, anyone can run that type of research. Being skilled isn’t enough anymore. The only investors worth hiring are those who can outsmart the machine.

Everyone gets the same answer now

I called AI a consensus machine in AI for Investment Research. It tells you what people think about a company. It does this based on the training data it learned from. A couple of years ago, getting that kind of read was tough. You had to read the filings, build a model, and understand the business to form an opinion. Now you can hand a 10-K to a chatbot and have most of it done in twenty minutes. If you had never done it before, that’s a real leap. You go from knowing nothing to a solid, middle-of-the-road grasp of a company over lunch.

The catch is that everyone else’s chatbot produces the same solid, middle-of-the-road grasp. And it’s worse than common because the answers converge. Doshi and Hauser, researchers in AI and creativity, asked people to write short stories. Some got help from a model, while others wrote on their own. The AI stories did better. This occurred as a result of the improvement in the weaker writers. The strongest writers saw no change at all. But the AI stories also started to resemble one another. Each writer did better work, and the work blurred together.

That convergence is what matters for investing. Now, anyone can write a decent thesis on a stock. But they all end up with similar ones. This happens because they use the same model, which is based on the same consensus.

How to tell if your edge is real

Back when a good thesis was rare, having one was worth a lot. A decent thesis is now free, and they all sound the same. So, decent work has little value. People will only pay for a unique, correct viewpoint.

Check the stocks you chose. Go through the list and ask if a chatbot would have suggested buying each one. Anywhere it would have said yes, you’re holding the consensus, and the consensus is free now. The picks that matter are the ones a chatbot would have flagged. It’s the stock nobody wanted. You held it for a year and a half, convinced the crowd was wrong.

If the honest answer is “a chatbot would have said buy” every time, move that money to index funds. You’re not giving anything up because the analysis you were doing is free to everyone now anyway.

I’ve used AI heavily for investment research for close to a year now, across all the major models. I wrote about the workflow that’s emerged in AI for Investment Research, and most of it still holds. There are three specific failures I’ve run into enough times to name.

It’s not a calculator

A calculator does math. An LLM generates sequences that look like the output of a calculation. Sometimes the sequence is correct because the training data contained the right answer to a similar problem. Sometimes it’s correct because the model walked through the reasoning in a way that happened to produce the right number. And sometimes the reasoning sounds right and a digit moves anyway.

I’ve caught this on dilution math, on weighted averages, on CAGR calculations, on discounted cash flow models. In every case the model set up the problem correctly. The variables were defined, the formula was right, the steps were ordered the way I would have ordered them. The error was somewhere in the middle of the arithmetic, and nothing about the surrounding output flagged it.

This is what makes it dangerous. A junior analyst who got the math wrong would also probably get the setup wrong, or hedge their answer, or flag uncertainty. The model does none of that. It produces the confident, polished output of someone who has checked their work, except it hasn’t.

So I use AI to structure the analysis and lay out which variables I need, and then I do the math in a spreadsheet. Any number that’s going to inform a decision gets recomputed outside the model.

It will use information that isn’t true

The second failure mode is harder to catch because it doesn’t have the clean tell of bad arithmetic. The model will confidently use information that is outdated, partially correct, or pulled from a context where it doesn’t apply.

Ask one of these tools about a company’s most recent quarter and there’s a real chance it gives you the quarter from two years ago, or mixes up two segments, or cites a revenue number that was correct on a different reporting basis. The training data has a cutoff, the web search results are noisy, and the model doesn’t distinguish well between “this fact was true in 2023” and “this fact is true now.” It treats them with the same confidence.

What’s changed how I use these tools is that I almost never let them work from their training data anymore for anything where the answer depends on what’s true right now. If I want analysis of a quarter, I upload the 10-Q. If I want help on an earnings call, I paste the transcript. If I want to understand a competitor matrix, I feed it the filings. The model is excellent at synthesizing across documents you’ve given it. It’s much worse at retrieving the right documents on its own, and worst of all at pretending it has retrieved them when it’s actually filling in gaps from training data.

This is the inversion of how most people use these tools. Most people ask a question and let the model figure out where to get the answer. For investment research that’s the wrong direction. You have to do the retrieval. The model does the synthesis.

It doesn’t know what’s material

A good analyst reading a 10-K knows that one paragraph in the MD&A about a specific customer concentration matters more than four pages of risk factors that every company in the sector includes verbatim. She knows the auditor’s report is boilerplate ninety-nine percent of the time and that the one time it isn’t, that’s the most important page in the document. She knows the segment disclosures often tell you more than the headline numbers.

The model doesn’t have any of that. It treats every sentence in a filing as roughly equally weighted, because that’s how text-based models read documents. When you ask it to summarize a 10-K, it gives you a competent overview that hits the major sections and misses the specific lines that would actually move your thesis. The summary is correct. It just isn’t useful, because materiality isn’t a property of the text. It’s a property of what an experienced reader knows to look for.

This is the place where the gap between AI and a trained analyst is largest, and it’s also the place that’s hardest to see from the outside. The output looks like analysis. It has the texture of analysis. What it doesn’t have is judgment about which facts in the document deserve attention, and that judgment is most of what a good analyst is being paid for.

The way I use AI now reflects this. I treat it as a powerful search tool for finding information across documents, not as an analyst for telling me what the information means. If I ask it whether a company has mentioned a specific topic in the last four quarters, the answer is fast and reliable. If I ask it what’s most important in the most recent quarter, the answer is generic and shaped like every other answer. The first question is a retrieval problem and the model is great at retrieval. The second is a judgment problem and the model has no judgment to apply.

The shared failure mode

By the time you notice the model was wrong, you’ve already used it to make a decision. That’s true for the arithmetic, the outdated facts, and the missed materiality, and it’s the reason all three matter more than they would in a domain with slower feedback loops.

The time savings is real. The week I used to spend reading filings and building competitor matrices is now a couple of hours. The synthesis across documents I’ve uploaded is excellent. The drafting of bull and bear cases as a starting point is useful. The arithmetic, the currency of the information, and the judgment about what matters are still on me.

Anyone selling you AI as a complete research tool is either not using it on real positions or hoping you don’t notice when the model is wrong.

The trade doesn’t change my core philosophy. The vast majority of my portfolio sits in low-cost index funds running on autopilot, and that boring system is the reason my financial life isn’t dependent on any individual pick going right.

In I Analyzed Stocks for a Living, I wrote about why even professionals with Bloomberg terminals struggle to beat the market. In Five Investing Beliefs That Sound Smart but Cost You Money, I called “do enough research and you can pick winners” one of the most expensive myths in individual investing.

Both of those things are true, and I still pick stocks anyway. The slice of my portfolio I run actively has produced enough good outcomes that the occasional bad one hasn’t mattered, and I think there’s value in walking through specific trades to show what the framework actually looks like in practice.

How I found it

Applied Blockchain in mid-2022 was a tiny company providing data center hosting for cryptocurrency miners. The stock traded around a dollar. The Fed had raised rates seven times that year, crypto was in a brutal winter, and Ethereum was months away from shifting to proof-of-stake, which would eliminate the economic basis for the kind of mining the company supported. By every conventional measure, you didn’t want to own this stock.

What caught my attention was the gap between what the stock priced in and what the company actually owned. The market was valuing Applied Blockchain as a crypto mining company in terminal decline, which is fair enough, because that’s what it was. But strip out the crypto business and what remained were data centers. Buildings in specific locations, with power contracts and cooling systems already operating, that had a use case completely independent of whether anyone ever mined another Bitcoin.

I had a thesis about what those data centers would be worth.

The thesis

In mid-2022, demand for high-performance computing was already accelerating, and the constraint wasn’t going to be silicon. It was going to be the physical layer underneath. Power capacity, cooling, the ability to actually plug something in at scale. You can’t just build a data center overnight. The permitting, the grid connection, the construction itself, all of it takes years. Anyone who already had operating capacity in 2022 owned something the market would eventually have to pay up for.

I didn’t know which specific use case would drive the demand. AI was an obvious candidate, but it could have been general cloud computing expansion, or scientific computing, or any number of other things. The point wasn’t to predict the application. It was that compute demand was a one-way bet, and the physical infrastructure to serve it was a constrained resource.

Applied Blockchain owned that infrastructure. The market was treating it as worthless because of what it was currently being used for. That was the gap.

My thesis was that the assets were worth multiples of where the market priced them, and that the path to realizing that value would come either from a strategic pivot, an acquisition by someone who wanted the capacity, or a re-rating as compute demand made the underlying assets visible. The asymmetry was clean. Big upside if any of those paths played out, and a downside floor set by the value of physical infrastructure that existed regardless of what happened to crypto.

What happened

The pivot happened in May 2023. The company (now renamed Applied Digital) launched specialized AI cloud services and announced its first major contract worth up to $180 million. The stock jumped 25% that week. Over the next two years they kept executing: a $5 billion leasing agreement with a hyperscaler, a direct investment from Nvidia, revenue growth of 250% year over year by fiscal Q2 2026. The stock went from around $1 to over $40 at its peak.

Why I sold

I sold in January 2026 at $38.61 because the stock had run too far, too fast. After a 34x move, it seemed more likely to go down than to keep going up. I took the profits and redirected the money to positions where I still saw a gap.

The trap with winners this big is that you start to feel like you have a special read on the company and should hold because you saw it first. The price doesn’t care that you bought at $1. It only cares about the next dollar of value the company creates, and at $38 the market was already paying for several years of that value in advance. The original trade was the gap between $1 and what the assets were actually worth. Holding past $38 is a different trade, on a different setup, and one I wouldn’t have opened from scratch.

How I think about the individual stock slice

A meaningful portion of my brokerage account, around 40%, is in individual stocks. The rest is in ETFs that provide the stable, diversified base. The ETFs are what make it possible to take real risk on the individual names without my financial life depending on any one of them.

This is the part of the personal finance internet that doesn’t get talked about much, because most of the content is calibrated for people who haven’t built the foundation yet. The advice for someone with no emergency fund and no automated investing is correct: stay away from individual stocks, just buy the index. The advice for someone who already has the foundation looks different. Once the base is solid, you can introduce asymmetric exposure on top of it, and that’s where the actual wealth creation happens for individual investors. I wrote about this framework in The Asymmetric Bets Framework, and the individual stock slice is one of the places that framework actually applies in my own life.

Within that slice, the principle I keep coming back to is buying when the consensus is against you. This part is uncomfortable by construction. If the trade felt obvious, the price would already reflect it. Most of the positions I’ve taken that worked involved buying something other people were selling. So did most of the positions I’ve taken that didn’t work, which is the part nobody mentions when they tell these stories. Contrarianism on its own isn’t a thesis, it’s a precondition for one.

One more thing

I bought Applied Blockchain in July 2022, four months before ChatGPT launched. The thesis required looking at a dying crypto miner and seeing a compute infrastructure play that the market wouldn’t be willing to price in for another year or two.

I’ve spent the last year using AI heavily for investment research, and I’m confident it would not have flagged APLD as a buy at $1 in 2022. The consensus view at the time was that the company was in terminal decline, and consensus is what AI returns when you ask it for an analysis. I wrote about this in AI for Investment Research, but the APLD trade is the cleanest example I have of why it matters. The setups that produce returns like this one live in the gap between what the public information looks like and what’s actually going to be true, and that gap is where AI is structurally weakest.

The honest answer, after a year of using these tools across several different models on real positions in my portfolio, is neither.

I moved from ChatGPT to Grok in the fall because Grok handled financial questions better. Then I moved to Claude when the more advanced model came out and the analysis got a noticeable jump in depth. I tried Perplexity along the way and wasn’t impressed. The tools have improved fast enough that anyone writing about them with confidence is going to look a little silly in six months, including me. So take this as a snapshot from May 2026 rather than a final word.

AI has made me a better investor, just not in any of the ways it was sold to me. The pitch from finance Twitter is that AI will democratize access to institutional-grade research and let retail investors find the next ten-bagger from their couch.

What AI is actually good at

When I was in equity research, the first week of covering a new company was almost entirely mechanical. You read the last four 10-Qs and 10-Ks, pull every press release for two years, listen to the last eight earnings calls, and build out a basic financial model alongside a rough competitor matrix. None of this required judgment. It required time, and it produced the substrate on which the actual thinking could happen.

AI does roughly that week of work in a couple of hours. You can have it read the last four quarters of a company’s filings and surface the major themes management has been emphasizing, pull a clean list of every competitor mentioned across those filings, or generate a serviceable version of the bull and bear case as presented by sell-side analysts who cover the name. None of these outputs is differentiated, but each one would have been a half-day of work ten years ago and is five minutes now.

Sentiment scanning is the other thing it handles well. What are people saying on the company subreddit, in the StockTwits feed, in the comments under recent YouTube earnings recaps? You can get a reasonable read in a few prompts. None of that is alpha by itself, but it’s data you couldn’t easily access before, and occasionally it surfaces something useful.

What I use it for most is getting oriented on complicated situations. A biotech running three trials at once with a patent dispute on top of it, or a roll-up that’s grown through 14 acquisitions in four years, can have a volume of disclosure that takes a full weekend just to read. AI gets me through that in an hour, and what would have been a Saturday is now the first hour of looking at a company.

What AI is bad at, and why it matters

The way you actually make money in individual stocks is by knowing or believing something about a company that the market has not yet figured out or accepted. That gap, between what the price reflects and what is actually true, is where alpha lives. Every big winner I’ve had, and every loser, came from a mismatch like that. The market eventually figured it out and the stock moved.

AI can’t find those mismatches.

Trained on public information, the model gives you back a tidy version of the consensus view. The bull case it constructs is the same one already reflected in the price, along with the bear case and the standard list of risks the sell-side has been flagging for months.

This is the opposite of what you need to find a mispriced stock. You need a non-consensus view, supported by reasoning the market hasn’t fully absorbed, and AI is structurally a consensus machine. It cannot give you an edge because the consensus view is, by definition, not an edge.

The other thing AI doesn’t do is hold a contrarian position under pressure. Some of the best calls I’ve made required sitting with a thesis for 18 months. The stock went nowhere, smart people told me I was wrong, and I kept asking myself whether I was the idiot. AI has no position to defend. It revises its view based on whatever’s in the context window. Ask it the same question three times with slightly different framings and you’ll get three different answers.

AI is bad at math. And not in a way that’s easy to catch.

It sets the problem up correctly and walks through the logic in the right order. Then somewhere in the middle of a DCF or a CAGR calculation, it drops a digit, and the final number is wrong. I’ve caught it doing this on dilution math, on growth rates, on weighted averages. The prose around the error is always confident, which is the part that makes it dangerous. I now run every number AI gives me through a calculator before I use it for anything.

The workflow that actually works

The way I use AI now isn’t to replace research. It’s to compress the time it takes to do the parts of research that don’t require judgment, so I have more time for the parts that do.

For any company I’m looking at seriously, I run a version of the same loop.

I ask one model for the strongest bull case it can construct, with specific numbers and timeline. Then I ask the same model for the bear case with the same level of specificity. I take both and feed them to a different model for critique, and then I feed the critiques back to the first model and ask what its strongest response would be. By the end of a couple of cycles I have a sharper picture than I started with, and more importantly, I have a sense of where the models are uncertain.

That sense of where the models are uncertain is the thing I’ve started to trust. If I ask three models for a real bear case on a stock and none of them can produce one that goes beyond generic risks, that’s a signal. The downside might actually be limited, because the bear case isn’t sitting in the public information set in any organized way. Conversely, if every model gives me the same coherent bull case in slightly different wording, I should assume that case is already priced in, and my edge has to come from somewhere else.

I don’t trust AI’s recommendations. I trust the texture of its disagreements with itself.

The other thing I do is calibrate its risk-return suggestions against my own profile. If you ask AI for “stocks with huge growth potential,” it will hand you a list of preclinical biotechs and pre-revenue lithium miners that could plausibly return 10x and are statistically much more likely to return zero. The model isn’t wrong, exactly. Those stocks do have huge growth potential. The model just has no way of knowing that “huge growth potential” for you means asymmetric upside with bounded downside, not lottery tickets. Risk and return are a personal question, and asking AI to optimize for them in isolation will produce a portfolio you should not own.

Where this is going

The biggest mistake individual investors are about to make is assuming AI will close the gap between them and institutions. It’s widening it.

Consumer AI gives you the equivalent of a smart junior analyst who reads filings fast. Inside hedge funds, AI is being integrated with proprietary data sets, real-time market feeds, and channel-check workflows built over decades. A hedge fund using AI on its own data is doing something different than you using AI to summarize a 10-K. The information asymmetry I wrote about in I Pre-Wrote My Research Reports Before the Earnings Call Even Happened has been augmented on both sides, but more on theirs than yours.

The floor has moved. Someone running a real research loop with these tools will make better stock decisions than someone picking based on whatever’s trending on FinTok. The ceiling hasn’t moved at all.

AI is a useful research associate and a useless portfolio manager. It can help you understand a company in a fraction of the time it used to take, but it won’t tell you whether to buy it, and any tool or prompt that claims otherwise is selling you something. The judgment is still yours. So is the conviction to sit on a thesis for 18 months while everyone tells you you’re wrong, which is where most of the returns actually come from.