Blind Extrapolation as a Powerful Force in Finance

In this issue:

• Blind Extrapolation as a Powerful Force in Finance—You can solve many financial mysteries if you assume that there's a substantial cohort of investors who form their predictions about returns mostly by looking backward.
• Midjourney Medical—In some parts of AI, saying that a process will start to get interesting with a few orders of magnitude more data is bullish, but depending on how the data gets collected and what happens next, it can be bearish.
• Chips—Google copies a model that works.
• Financial Engineering—Reverse-acquihires haven't had time to build the precedents other deals have.
• Rigged Games—Fake bets, real ads.
• Reality is the Complement—If the cost of an AI-generated image is basically zero, the value of knowing an image is real goes up.

Talk to this post on Read.Haus.

Blind Extrapolation as a Powerful Force in Finance

A few weeks ago, The Diff wrote about building a market simulation to see if simple rules could produce complex emergent properties like value and momentum strategies producing excess returns. There must be something in the air (specifically: agentic coding dramatically speeds up the loop of writing a simulation, running it, looking at what’s weird in the results, tweaking it, and running it again). Here’s another team, Victor Haghani and Rich Dewey, with a look at their market simulation.

Earlier this year we gave a two-part talk at the Numercon Conference titled Who Killed The Random Walk?, which is based on ideas in a recently released research paper of the same title. The idea was to explore how agents with different expectations, beliefs, and constraints interact with each other to shape prices in financial markets. We let them trade with each other over decades of simulated time.

What emerged was realistic and illuminating. The interaction of these simple agents produced excess volatility, bubbles and crashes, fat-tailed returns, and persistent trends—all phenomena that have puzzled market observers for centuries. These are not merely theoretical curiosities. In 2007, quantitative equity funds suffered sudden, correlated losses when crowded strategies unwound simultaneously—one set of agents' forced selling triggering another's. During the 2021 meme-stock episode, coordinated retail options buying generated gamma squeezes that forced market makers into destabilizing hedging flows. In 2025, the "Rolling Thunder" drawdown ($, FT) revealed how quickly algorithmic strategies can cascade when they collide.

In this model, one of the magic ingredients was the inclusion of extrapolators, who naïvely assume that the market’s return is higher if it’s been high recently, and vice-versa. This is not quite the same thing as momentum, which doesn’t make a claim that future returns will be similar to the recent past, only that on average excess returns persist for a bit. The extrapolators tend to participate more in the market when it’s gone up recently, and then to put more of their money in cash when the market has declined. So they tend to keep both trends moving.

At one level, this is just directly adding some excess volatility and time-varying risk premia to the model in a direct way, by telling investors: have your equity risk premium be whatever the most recent realized equity risk premium was, even if that is far outside of historical norms. On the other hand, that’s what the literature says many investors do! Sometimes, it really is that easy.^[1]

We think this research is important to financial markets for two reasons. First, the character of markets is changing rapidly. New instruments such as zero-day to expiration options, perpetual futures and leveraged and inverse ETFs (often based on a single stock or commodity) have grown in popularity. Prediction markets and tokenized assets are likewise adding new dimensions to markets. Second, the composition of market participants is changing, with retail traders—whose actions can become correlated whether by design or contagion—and quantitative trading firms both growing in size and influence.

Modeling agent behavior can help us understand market dynamics and hopefully make them more efficient and less prone to crises. We don’t have good historical precedents for how large-scale use of zero-day options interacts with extreme volatility. We have some suggestive examples, like looking at the rollercoaster around Liberation Day. But that illustrates why this is so hard: how do you tease out what was Liberation Day-specific from what was caused by options? And if you’re predicting the next period of extreme volatility, how do you decide whether market participants repeat their mistakes or make equal and opposite ones instead?

For active market participants, identifying tipping points is crucial not just for profits, but for survival. Some firms involved in the episodes described above did not survive, suffered continuity-threatening losses, or, most embarrassingly, joined the category of asset managers whose compounded returns are decent and whose total lifetime alpha in dollars is negative because they were managing so much at the peak.

The importance of this research goes beyond financial markets as it might shed light on what Nvidia CEO Jensen Huang recently remarked will be the coming “age of Agentic AI.” The explosive growth of tools such as OpenClaw is suggestive of a future where agents take on many responsibilities for us, for example helping with healthcare decisions, financial planning and organizing our travel and entertainment activities.The full implications are uncertain, but we believe financial markets offer a particularly instructive lens for observation and understanding.

Financial markets possess three properties that make them an unusually good laboratory for studying multi-agent systems: clearly defined rules, powerful incentives (money, reputation, survival), and lots of data. Add to that relentless selection pressure—agents that consistently underperform are diminished, and those that outperform gain influence. Markets even capture subtleties like the difference between good outcomes on average and good outcomes for the average person. And they tend to incubate surprisingly complex emergent behaviors that can then be checked for realism against the real world: we can generate all the synthetic data we want and check it for realism against historical data that goes back centuries.

Researchers working at the intersection of finance and reinforcement learning have already shown that market-making algorithms, when left to learn and adapt, can converge on collusive equilibria—improving their own profits at the expense of everyone else, without any explicit coordination. If that can happen in a regulated, transparent market, what happens when AI agents negotiate prices, allocate resources, or manage logistics in less structured environments? At the Santa Fe institute collaborations between behavioral economists and physicists including Doyne Farmer, Brian Arthur and Blake LeBaron have examined agent interactions and how irrationality can cascade into socio-economic systems.

The frontier AI labs are also wrestling with these problems in an attempt to align increasingly powerful LLMs with humanity’s best interests and their designers’ stated intentions. The work on alignment and interpretability has received increasing focus and made encouraging progress. However, financial market episodes suggest that even thoughtfully designed systems and strategies can encounter unexpected outcomes in complex and dynamic systems—though they also suggest that a population of heterogeneous agents can achieve a sort of rock-paper-scissors equilibrium where one of them never quite dominates.^[2]

We believe these lines of inquiry all have merit and can help develop our understanding of agents acting in noisy and competitive systems. We think this is critical given widespread expectations that algorithmically driven agents will soon take on work previously done by humans; we’re at a brief moment in history where fairly autonomous software agents exist, but don’t yet outnumber humans by orders of magnitude. (That sounds implausible, but there was a similar point in the industrial revolution, when you could have observed that hydrocarbons are powering useful work and will clearly do a lot more of it than humans ever did. The amount of energy used to move an average car the distance of an average commute is about 10,000 calories, or about four full days’ worth of human energy consumption)

Financial markets have been confusing since José de la Vega wrote the first book about stock trading in 1688. Our simulations suggest the confusion is not noise—it is the predictable consequence of diverse agents pursuing different goals with different beliefs. Understanding those interactions will not eliminate the confusion, but it can transform it from a source of anxiety into a source of opportunity to better design and manage mission-critical systems.

Richard Dewey is CEO of Allometry Labs and the Managing Partner of Revenant Ventures. Victor Haghani is the founder and CIO of Elm Wealth and co-author of "The Missing Billionaires: A Guide to Better Financial Decisions." The paper, "Who Killed the Random Walk?", is available on SSRN.

1. Interestingly enough, you can get some of these results without extrapolators, if your market has leverage, trades have realistic price impacts, and at least some levered participants are valuation-insensitive, then on average they’ll behave like extrapolators because every price move in their favor gives them the collateral to make another trade in the same direction, pushing prices further in that direction. If prices move in the wrong direction, they get margin calls, which have the same effect even faster. But even though leverage is a mechanism that produces extrapolations, it’s not the only one, and actual extrapolation can also show up without it. ↩︎

2. One way for this to happen, for example, is for the lab with the most capabilities to impose the most safeguards, and for the next-best lab to compete partly by letting users do more, just with models capable of a little bit less. In that model, the leading lab will tend to be used to build detection tools and countermeasures to what the second-place one does. This is of a similar shape to the situation with Mythos right now. ↩︎

Diff Jobs

Companies in the Diff network are actively looking for talent. See a sampling of current open roles below:

• Lightspeed-backed team building the engineering services firm of the future is looking for founding members of technical staff excited about working alongside civil engineers to translate their domain expertise into the operating system that powers the next era of great American infrastructure. If you’re an engineer with strong product intuition, who's energized by access to users, and excited by the prospect of transforming how we design and construct our built world with frontier AI, this is for you. (NYC, SF or Remote)
• Ex-Palantir, Citadel founders building the meta-harness (the system that knows what hills to climb, and what the right loss functions are) for all the lucrative agents need full-stack engineers that understand that turning AI into economically valuable solutions means a system that includes deterministic infrastructure. If you’re curious about building a system that finds what the efficient frontier between determinism and stochasticity actually is, this is for you. (NYC)
• Series-A defense tech company that’s redefining logistics superiority with AI is looking for a MLE to build and deploy models that eliminate weeks of Excel work for the Special Forces. If you want to turn complex logistics systems into parametric models, fit them using Bayesian inference, and optimize logistics decision-making with gradient descent, this is for you. Python, PyTorch/TensorFlow, MLOps (Kubernetes, MLflow), and cloud infrastructure experience preferred. (Salt Lake City or NYC)
• A well-funded, Series C startup building the platform and agent primitives to drive operational transformation at large, complex institutions (starting with higher education) is hiring platform engineers. The work spans distributed systems, applied AI, and full-stack infrastructure, focused on deploying reliable agents that meaningfully bend institutional cost curves. (Remote)
• A hyper-growth startup that’s turning the fastest growing unicorns’ sales and marketing data into revenue (driven $XXXM incremental customer revenue the last year alone) is looking for a senior/staff-level software engineer with a track record of building large, performant distributed systems and owning customer delivery at high velocity. Experience with AI agents, orchestration frameworks, and contributing to open source AI a plus. (NYC)

Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up.

If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority.

And: we're now actively deploying capital into early-stage companies through Anomaly. Our focus is on defense, logistics, robotics, and energy. If you'd like to chat, please reach out.

Elsewhere

Midjourney Medical

Late last week, Midjourney, better known for image generation, announced that they're getting into the medical business, specifically whole-body 3D ultrasound tomography (a different kind of image generation!). Their goal is to make these scans ubiquitous, cheap, convenient, and, based on their render of the Midjourney Spa, perhaps luxurious. Doctors are not so enthused; Scott Alexander has a guarded take, and plenty of them on Twitter are pretty annoyed. Here is a representative case, though that doctor also walked it back. The key concern is: if we usually scan people who have symptoms, and we start doing precautionary scans, then people will request biopsies, which risk medical complications. And even if they don't, they'll be worried sick over what doctors call an incidentaloma, i.e. a growth that you weren't looking for, that doesn't have any symptoms, and that is almost certainly benign. The pro-Midjourney argument is: why are you treating the conditional probabilities as the same under different conditions!? Maybe there are some particularly pushy patients, but at some point a doctor is either a trustworthy medical professional, who will tell people no, or a doctor is a service provider who is selling to willing customers who will at least say they're comfortable with the risk they're taking. The more interesting outcome is that if scans get cheap and lots of asymptomatic people keep doing them, we'll get a rich set of historical data that will eventually be great training data once some of them develop illnesses. Now is just a really bad time to bet against anyone who is trying to capture snapshots of data and then predict the contents of the next snapshot.

The less-spoken subtext of this is that programmers and doctors shouldn't be fighting with each other when they could be going after their common enemy, the lawyers. Medical malpractice risk warps doctors' incentives, particularly around questions that are probabilistic when they're posed but that eventually have a definite answer. Which makes the doctors' situation a lot more sympathetic. For one thing, they don't want a bunch of people showing them scans and exaggerating the risk, then asking for treatment. If they treat, there's a risk that something goes wrong with the biopsy. If they don't, eventually one of their patients really will turn out to have cancer, and will be able to point to the scan and show—only with the benefit of hindsight!—that they gave their doctor a picture of it. And, at another level: they can't say this, either! If they say "I don't want my patients to take tests lest they get a superfluous operation, something goes wrong, and they sue me," that statement can get read back to them in court when they're being sued because they didn't order a test. So doctors have an unfortunate incentive to 1) resist any situation where some kind of diagnostic was previously run on sick people and is now being run on healthy ones, and 2) to be a little vague about why. The high-handed attitude that anyone who Googles a symptom or tries to learn about their own body is voiding their warranty is actually the next-best option.

But Midjourney is not the only company that's trying to use AI for medical research, and we'll probably see many similarly-shaped problems. To someone with a lot of computer hardware or a penchant for deep learning, every problem is easiest when you turn it into a data problem, and turning a healthcare problem into a data problem often means running a test on someone who isn't sick, and then telling them something that probably makes a lot of sense to a professional poker player or an ML engineer but that will be easy for a regular person to misinterpret. Doctors are obviously not afraid of new technology, and are actually surprisingly early and aggressive users of AI.

Chips

A rough guide to the chip business is that marginal cost goes down when fixed cost goes up, which 1) leads to some exciting swings in profitability if your new, low-cost capacity comes on line right as demand starts falling, and 2) means that a chip business can be more profitable if it can find more ways to guarantee demand. Google is taking some lessons from Nvidia in that respect ($, WSJ), by backstopping datacenters that use their TPUs. One way or another, Google captures upside from a lower marginal cost for TPUs: it can sell them through Google cloud, offer strategic partnerships, give big partners an alternative to Nvidia (or to Amazon), and, as a backup, use them in its own search, YouTube, etc. business. For now, many of Google's capital-allocation decisions are downstream from the fact that it's its own reserve bidder for inference.

Disclosure: long NVDA, GOOGL, AMZN.

Financial Engineering

A few years ago, some clever deal types realized that instead of going through the arduous process of acquiring a company, they could offer a hybrid deal where they paid off investors (partly), hired the team (also partly), and left behind some well-funded but mostly inert remnants. But a look back at the historical record for these deals shows that often, the acquired team doesn't stick around long. When there's a new financial maneuver, it can unlock some transactions that weren't possible before. But it also hasn't had time to accrete the kinds of stipulations, slow payouts, representations, etc. that traditional deals do.

Rigged Games

A pretty obvious way to market a gambling service is to show a video of someone placing a bet and winning. It's the rare case where a pure financial service can actually be pitched in a pleasantly direct way. However, this leads to temptation: make a video that isn't quite identified as an ad, and that shows a winning bet by using a dummy version of the site in question. That's what Polymarket appears to be doing ($, WSJ): they've paid online influencers to post clips that portray winning bets, but at least some of these bets were made on "poiymarket.com" instead, and referenced events where the resolution would have been different had the user actually made the bet in question. This is just incredibly sloppy on Polymarket's part: they've created a huge surface area and deliberately posted misleading news snippets on their branded marketing channel. It would be very hard for them to get away with this kind of thing.

Reality is the Complement

Getty Images shares have lost 95% of their value during the company's stint as a public company, at least partly because generative AI is such a great solution for the I-need-a-picture-of-something-here issue that leads to so much demand for stock photos. But that also means that even though every article with an AI-generated stock image has a customized illustration, they all tend to look the same. Which means that genuine photographs are, in a sense, more valuable than they were before because they signal that the content in question is not being mass-produced at minimal cost. And how better to distribute such images than through the same tools that make the substitutes? So Getty Images is worth twice what it was last week after announcing a licensing deal with OpenAI. There are more companies that fit this template, where AI is in one sense a substitute but in another sense turns what they compete with into a premium product by default.