[Week 47 of 2025] Vectors and Veils

Welcome back to Price and Prejudice with a few musings from Week 47 of 2025.

Embeddings Everywhere

Here's an interesting paper called "Asset Embeddings" which I had the pleasure of discussing at a conference at NYU this past Friday. The idea starts with something simple: in AI, an embedding is a way of turning a complex object (like a word or an image) into a set of numbers that capture its meaning and relationships. This paper applies that logic to finance. Instead of treating assets as isolated datapoints, it shows that investors’ portfolio holdings contain rich information about how assets relate to each other. And with tools from AI, we can learn these relationships directly from who holds what, without needing every accounting number or textual description that investors rely on.

What makes the paper interesting is not the abstract idea of embeddings, but the fact that the authors turn this idea into something operational for finance. By learning from the choices investors actually make, rather than the variables academics think matter, you get models that explain a lot more of what we see in markets. From an academic perspective, this is a welcome response to the critic that academics are drifting further and further apart from the complex reality of financial markets.

And indeed, the paper nudges the academic world toward a more flexible mindset. For decades, academic research has usually started by specifying a theoretical investor — with preferences, beliefs, and constraints — and then building models to match the data. The embeddings approach flips that order: learn the structure directly from the data first, then ask what patterns of preferences or beliefs might rationalize it. This doesn’t replace economic theory, but it does broaden the toolkit and helps the field escape some long-standing bottlenecks. It's very possible that this data-first, representation-focused approach will define much of the frontier in empirical research in finance for the next few years, even if the ultimate goal remains the same: tying prices back to human behavior and economic fundamentals.

What 13Fs Actually Reveal

Coming off the discussion of asset embeddings, it’s worth remembering where the raw material for many of these models actually comes from. In U.S. equity markets, the closest thing we have to a standardized snapshot of institutional positions is the 13F filing — a quarterly disclosure required of large investment managers, listing their U.S.-listed equity long positions. It is, in theory, a window into how the big players are allocated. In practice, a 13F shows only one side of the book (the longs), excludes things that may be potentially important (shorts, swaps, futures, credit, funding trades), and is released with a lag long enough for a significant chunk of the economic exposure to have mutated.

This article goes one step further to argue that the misunderstanding isn’t just about what’s missing, but about how deeply misleading the visible parts can be. The article points out that even when a 13F appears to show something meaningful — a large position, a sector tilt, a sudden stake in a household-name company — that line item often has almost nothing to do with the manager’s true economic bet. What looks like a conviction position might really be the anchor leg of a pair trade, the carrier for a volatility structure, or the placeholder around which a much larger, much more intricate strategy is slowly brewing. Viewed through this lens, reading a 13F in isolation is like reading a novel by flipping to every 20th page: you see words, even plot points, but miss the connective tissue that makes the story coherent.