[Week 27, 2025] Overcast and Overfit

Welcome back to Price and Prejudice with a few musings from Week 27 of 2025.

Fair Weather

The connections between finance and weather have a surprisingly long history. The first marine insurance contracts in medieval times covered shipwrecks due to storms, establishing the precedent that weather risk could be transferred for a price. Weather derivatives are also a multi-billion dollar market where companies hedge against rainfall, temperature, or wind speed. Hedge funds are continuing the tradition of hiring meteorologists, and some financial products even bear meteorological names like Ray Dalio's famous "All Weather" portfolio.

Perhaps it's not a surprise that this article on the recent advancements in weather forecasting parallels that of the developments in the competition for alpha in the financial sector. The article highlights two main forces: better predictions through AI and better data through improved sensor networks. These are precisely the twin pillars that have always defined success in (quantitative) finance: proprietary model architecture for superior forecasts (e.g. the alleged use of hidden Markov models at Renaissance) and various sources of alternative data (e.g. satellite images of parking lots for forecasting revenue). Whether you're predicting storm paths or stock prices, the fundamental challenge remains the same: extracting signal from noise in complex, chaotic systems.

Banking on Tolkein

Incumbents in banking enjoy a crucial structural advantage: deposit storage depends heavily on perceived safety, and reputation derives largely from longevity – something only established institutions can demonstrate. This creates a natural moat that explains why community banks founded in the 1800s still compete effectively against fintech startups armed with superior technology. The solution for challengers is market segmentation, carving out niches that incumbents either ignore or find unprofitable to serve. This was Silicon Valley Bank (SVB)'s strategy before its spectacular collapse in 2023. SVB built a franchise around venture capital firms and tech startups that traditional banks viewed as too risky, until interest rate risk and deposit concentration destroyed the institution in days

This article describes how tech billionaires Palmer Luckey and Joe Lonsdale are launching Erebor, a new bank targeting the same innovation economy that SVB once served. What's particularly interesting is the naming strategy: Erebor references the "lonely mountain" in Tolkien's Lord of the Rings, whose treasures are reclaimed from a dragon. The literary allusion signals permanence and gravitas in a way that "Silicon Valley Bank" never could – geographic names suggest transience, while mythological references imply timelessness. It's a subtle but potentially important branding choice for an institution that needs to convince customers it won't suffer SVB's fate.

Virtue of Complexity

This article covers an active academic debate about whether bigger AI models make better stock pickers, one where I find myself with a front-row seat to both sides. Stefan Nagel, who provides one of the key critiques mentioned in the piece, was one of my main advisors at Chicago, while AQR is a firm where I worked briefly. Basically, the paper in question argues that complex machine learning models consistently outperform simpler ones in financial forecasting, contradicting the conventional wisdom about overfitting. Nagel's critique is that Kelly's complex models aren't actually learning sophisticated patterns, but they just happened to implement momentum strategies with recency bias, making their outperformance a statistical coincidence rather than genuine insight. (NB: Next week at a conference, the authors will have a chance to directly respond to each other's critique, which will also be livestreamed.)

Taking a step back, this represents one of many ongoing efforts to transplant core insights from AI/ML literature into finance. The challenge is that prediction in social sciences is fundamentally more difficult than in natural sciences: stock prices reflect human behavior and strategic interactions rather than physical laws. While deep learning achieves remarkable results in image recognition or language processing because the underlying patterns are relatively stable, financial markets are populated by agents who actively try to exploit any predictable patterns, potentially eliminating them. So while I wouldn't dismiss the possibility that larger models could eventually prove superior, I'm skeptical that results will transfer as mechanically and directly as proponents hope.