The Bitter Lesson Is The Final Competitive Moat

Introduction

Up until not too long ago, the modal way of preserving a moat in quantitative finance was to operate in absolute secrecy. This was in part because the traditional quant edge was a secret, a signal you found, that little to no one knew about, and was now yours to guard.

I strongly believe that the importance of “absolute secrecy” will rapidly drop in the coming times. In the long tail of AI commoditising intelligence, the final frontier of competitive advantage is some form of distribution and a business with an ability to scale in effectiveness on compute.

This article is mostly written from the perspective of a business in the field of quantitative finance, over numerous discussions about @openforage. However, I think the central argument applies to any firm whose talent (know-how, experience) and IP have traditionally been primarily the moat.

What Is A Moat?

For serious quantitative finance practitioners, alpha is the name of the game. It is a pool of returns that arises as a result of your skill, independent of any form of systematic sources of returns that are entangled with gross returns.

For the layman, alpha is the money you generate on top of the market. If you have positive alpha (are skillful), you should be able to generate returns in all market conditions. This is especially true if you can hedge out the systematic sources of returns that are NOT part of your alpha (e.g. markets going up or down).

Alpha is essentially zero-sum. A particular investment process in a particular market has a finite amount of alpha that can be expressed in fixed dollar amounts. If you have monopolistic control and generate $1,000,000 of alpha, and a competitor comes in and runs a similar investment process, his profits in alpha will directly come from a share of your alpha.

A moat is an advantage that allows you to keep competitors from “stealing” your alpha. An example of a simple moat is intellectual property / know-how. If your competitor does not know how to replicate your investment process, then he cannot compete with you in the same pool of alpha.

The Old Guard And Signals

For decades, the source of alpha was discovery in know-how, or “research,” as it is often called. Firms would spend tremendous amounts of money to hire the sharpest people, give them enormous amounts of data, and have them trawl through it to find predictors nobody else had.

Good predictors were worth a lot of money, as no one else had them, and they would have high signal-to-noise ratios. It is not uncommon to hear of stories of old where a handful of signals would carry entire quant trading teams/firms. For those in the know, there are behemoth firms whose lineage traces to essentially a handful of mean-reversion signals.

It didn’t help that computational costs were much higher then, so the number of signals that could be run economically in production was much lower. This made every signal that much more valuable, and a low number of signals implied that every incremental signal had to have sufficiently high signal-to-noise ratios.

When every signal could change the course of your firm, your moat was your secrecy. You would do everything you needed to do to prevent as few people as possible from finding out about your signals.

Building Things Is Hard

Even as compute got faster and cheaper, research moved towards increasing the number of signals that could be generated, or faster infrastructure, or more modular code that would allow you to compete around the globe.

Sophistication increased across the entire investment process. Risk management departments started formalizing “quant coverage”; execution systems got faster, better, more efficient. Central liquidity books became a thing, and netting became a competitive advantage.

Every firm doubled down on “technology”, and this was the modernization of quant finance. The firms that could do this best had the moat of being able to invent, wield, and manage “technology” better than their competitors. The largest behemoths built today are all beneficiaries of this paradigm.

Even as the industry moved towards modernization, the reasons for secrecy changed but the advantages did not. Firms were highly incentivised to keep everything on the down low, because knowing how to invent, build, and manage these technologies would invite competition and decrease alpha.

Things Are Changing

Everyone Knows Your Secrets

Today, things are a little different. I don’t think anyone with a sufficiently high bird’s eye view would disagree that strategies and IP have largely “converged”, and almost all “new discoveries” are marginal now.

This is largely because of the immense amount of information that has leaked from all of the top quantitative firms as a result of talent switching firms and just the general march of time, given anonymous accounts on fintwit, disappearing messages on Signal, etc.

Most signals with a name and an economic rationale are “known”. The value of secrecy is diminishing daily, and perhaps is now maximally useful only in knowing what signals are currently performing at a given point in time, rather than the signal itself, since there is so much arbitrage capital chasing all “known” signals that virtually all signals have degraded to having “pockets of predictability”.

When new signals are discovered, they are very often variants or combinations of older, established, “foundational” signals, rather than entirely new species.

There used to be some operational alpha in distilling research papers faster and more effectively so you could deploy signals earlier than your competitors. Everyone competent is virtually running an endless research pipeline consuming all new papers now.

So, secrecy over the known obviously cannot be a lasting competitive advantage.

Everyone Knows How

With LLMs and how good they have become at doing research and writing software, it is a matter of time before they can replicate all known concepts.

You may not know what a “central liquidity book” is, but there is a whitepaper out there that describes it in sufficient detail that an LLM can get to 90% fidelity in one try.

In this world and the coming one, there is no space for “know-how” as a moat.

The Last Frontiers

My beliefs for sustainable competitive advantages in a future where intelligence is increasingly commoditised can be roughly summarized in the following 3 parts.

Proprietary Exhaust & Positive Flywheels

If you can improve faster than you decay, you will survive the march of time. If your business can produce something unique and proprietary that allows you to improve as a result of having it, you will be able to compound and stay alive, and before long, out-compete your competitors.

An example of this is proprietary data as a result of your business operations. Imagine if you were a farmer, and you grew oranges, and as a result of growing oranges, you produced data that taught you how to grow oranges in 5% higher quantity that was unique to you and unavailable to your competitors. Before long, you would be the ultimate orange king, and corner the orange market.

(Maybe) Distribution?

I’m increasingly convinced that as AI has increased and continues to increase the amount of noise in the world, there are going to be those that will be drowned out by the noise, and those that aren’t.

Real markets are not frictionless, and there is a cost to discovery. There are going to be shittier products with amazing distribution that cuts through the noise, and there are going to be great products with virtually no distribution that will die irrelevant.

I’m not sure if this will continue to be the case as we relegate more and more of our decisions to AI? They are unlikely to have the same emotional appeal as us, and their cost structure for convenience is going to be vastly different from ours.

In THAT world, it may be the case that our agents will always seek and go for the best product in our favor — caveat: that they are actually able to find the best solution among the noise.

The Bitter Lesson

This is my personal favorite, and has been a belief of mine since I could reason about quantitative trading. The idea of the “bitter lesson”, in layman’s terms, is simply an observation that oftentimes, the best solution is to simply have something that scales with compute, and takes advantage of compute scaling laws.

I think one could very well apply this to businesses. I don’t think all businesses are applicable to that form of scaling. Further, I don’t think all investment processes are applicable to that form of scaling.

The litmus test is simple: can you find a relationship between the PnL you can generate and the amount of compute you have access to? For most investment processes, the relationship drops to 0 quickly. For others, it is often a weak relationship.

I strongly believe that the only kinds of investment processes that will survive the future are the ones where there is a clear and obvious relationship between PnL and compute. That is, you could run it like an AI lab, raise for compute, and use it to scale your investment process to increase capacity and performance accordingly to produce more PnL.

Over time, for businesses that can scale with compute, the infrastructural cost of maintaining such compute to meaningfully compete itself becomes the moat. That is, the cost to compete becomes so high that virtually no one else is able to meaningfully compete. We see this in the economics of frontier labs and already at the apex of quantitative finance in the most competitive markets.

If you take a gander at the top firms trading US equities, the amount of resources needed to meaningfully compete at the same scale and game is so large that the number of successful new entrants is ~0. I can’t imagine a world where this gap narrows rather than widens.

Putting It Together

Generally speaking, I think the name of the game is something along these lines: pay your dues and build infrastructure and an investment process that will allow you to scale with compute. Figure out how to design something that is amenable to producing proprietary exhaust and can create a positive flywheel from your existence.

Find a spot on the Pareto frontier where you can be the most sophisticated player, and if you do it right, you might just survive, day after day. Survive for long enough, and you might soon find that you are the last one standing.