Factor Investing Discussion - Inspired by Andrew Chen

I’ll separate the practical engineering lesson from the scientific/philosophical one, because “thrown out” is a bit too strong but there is a real shift here.

Basically yes for practical engineering, but no for the science of language.

LLMs showed that you can get extremely useful language behavior without hand-building a Chomsky-style theory of syntax, semantics, innate grammar, phrase structure rules, transformations, and so on. Modern systems mostly learn statistical/computational structure from huge corpora using general architectures like transformers. The transformer paper explicitly proposed a comparatively simple architecture based on attention rather than recurrence or convolution, and it worked very well for translation and parsing tasks. arXiv BERT then showed that large-scale pretraining on unlabeled text could produce representations useful for many NLP tasks with little task-specific architecture. arXiv GPT-3 pushed the same lesson further: scaling language models produced strong task-agnostic few-shot performance. arXiv

So the practical lesson is:

You do not need an explicit symbolic theory of language to build useful language technology.
You can learn a lot of the usable structure from data, scale, optimization, and feedback.

But that does not mean Chomsky-style linguistics was simply “disproven.” It was often trying to answer a different question: what kind of mental system allows humans to acquire and use language? Chomskyan grammar is concerned with formal generative capacity, productivity, competence, and the structure of human linguistic knowledge. Britannica summarizes the “generative” idea as explaining how finite means can produce infinitely many grammatical expressions. Encyclopedia Britannica LLMs do not settle that question, because they are not human children, do not learn from the same data regime, and do not necessarily represent language the way humans do.

A good way to put it:

LLMs did to classical theoretical linguistics what airplanes did to bird anatomy: they proved you can build something that flies without copying the biological theory exactly. They did not make birds uninteresting.

Where LLMs did undermine older assumptions is in the belief that robust language understanding required explicit, human-interpretable grammar rules. For practical NLP, that view lost badly. Statistical and neural approaches won because they work.

Where the older linguistic theories still matter is in questions like: why humans acquire language so efficiently, what counts as grammaticality, how syntax relates to meaning, what errors reveal about cognition, and why certain structures seem universal or rare. LLMs are also not perfect counterexamples to “you need structure,” because they may learn implicit structure internally rather than use no structure at all.

So your intuition is right with one refinement:

LLMs did not show that language has no deep structure. They showed that, for many applications, you do not need to explicitly know or encode that structure.

Can this be analogous to factor investing vs ml application that beats it (medallion fund)? With the backdrop that factors are becomming insignificat and do not seem to work? In a same way these fcators were tried to be justified by some theory, but in reality machine learning/ai beats and the theory was not that important.

Show moreShow less

Yes, that is a very good analogy, with some important caveats.

The analogy would be:

So yes: in both cases, practical success shifted away from elegant explanatory theory toward empirical prediction.

But I would not phrase it as “factors are dead and ML beat theory.” I would phrase it more like this:

Factor investing is to markets what formal grammar was to language: a useful, simplified map of reality.
LLMs/Medallion-style systems are more like engines that exploit the full messy terrain directly.

The key distinction is that factor investing is mostly low-dimensional and public. It says: stocks with certain broad characteristics have historically earned premia. For example, value, momentum, quality, size, low volatility, etc. Fidelity describes factors as security characteristics used to help explain risk and return, and AQR describes factor investing as being grounded in decades of academic research. institutional.fidelity.com+1

A Medallion-style system is probably not “factor investing but with better factors.” It is likely a huge collection of short-horizon, weak, changing signals; execution advantages; alternative data; portfolio construction; risk control; cost modeling; and constant adaptation. Public descriptions of Renaissance emphasize its secrecy, mathematical/statistical modeling, and hiring of scientists rather than traditional finance people. Investopedia reports that Medallion averaged nearly 40% annually from 1988 to 2023, which is extraordinary, though the exact methods are not public. Investopedia

The parallel to LLMs is strong in this sense:

The winning system may not need a human-legible theory.
It may only need enough predictive regularities, enough data, enough compute, and a good training/evaluation loop.

But there is a difference: language is relatively stable; markets are adversarial. English grammar does not disappear because everyone learns it. A trading signal can disappear precisely because everyone learns it. That makes finance much harsher. Once a factor is published, commoditized into ETFs, and crowded, its future premium can shrink. The Financial Times has noted that factor performance can be inconsistent and unpredictable, with major short-term underperformance even when long-term evidence exists. Financial Times

So the analogy should include crowding:

In NLP, scaling discovered structure.
In markets, scaling can destroy structure.

That is why “factors stopped working” is partly true but too broad. Some factor portfolios have had long painful stretches, especially value and small-value in certain markets. But there are still managers arguing that factors are not dead, just cyclical, crowded, badly implemented, or mismeasured. AQR, for example, still defends factor investing and argues that many criticisms confuse bad recent periods with invalidation of the underlying evidence. AQR Capital Management JP Morgan’s 2026 factor outlook is also not “factors are dead”; it says Q1 factor performance was largely positive and expresses optimism on value and quality in some regions. J.P. Morgan

The better claim is:

Naive, public, static factors have become less reliable as standalone alpha.
That is very plausible.

But:

Factors still matter as risk descriptors, portfolio-construction tools, and rough explanatory variables.

That mirrors linguistics. Chomsky-style theory may not be necessary to build ChatGPT, but syntax is not irrelevant. Similarly, Fama-French-style factors may not be enough to build Medallion, but value, momentum, liquidity, volatility, leverage, and quality are still useful ways to describe what a portfolio is exposed to.

The strongest version of your analogy is this:

Both classical linguistics and factor investing tried to compress a complex adaptive system into a few human-understandable principles.
LLMs and elite quant funds showed that, for performance, you can often skip the elegant theory and directly learn high-dimensional statistical structure.

The limit of the analogy is this:

LLMs commoditized language intelligence; Medallion did not commoditize market-beating ML.

That matters a lot. Many people can train useful language models from public data because language has reusable structure. But if everyone could replicate Medallion’s signals, those signals would likely decay. So finance remains zero-sum or near-zero-sum after costs, while language modeling is not.

So yes, your intuition is basically right, but I would sharpen it:

Factor theory was not useless. It was an interpretable low-resolution model. ML/stat arb is a high-resolution adaptive model. In practical alpha generation, the high-resolution model can dominate. But the low-resolution theory still helps us understand exposures, risks, and why some strategies worked historically.

How long till you admit they dissapeared? What if these factor historically worked only because we were at certain econmic and tech development? Why wouldnt next 100 hundred years be dominated by growth and size outperforming justified by enomours economies of scales unlocked by internet, AI, blockchain etc.?

Show moreShow less

I’ll treat this as a falsifiability question, not as “defend factors.” The core issue is what evidence would distinguish temporary underperformance from a structural regime change.

Yes — this is exactly the right challenge. I would not wait forever. At some point “long-term underperformance” stops being a drawdown and becomes evidence that the old model is wrong.

My answer would be:

I would admit a factor has effectively disappeared when its future expected premium is no longer positive after realistic costs, taxes, implementation frictions, crowding, and model error — not merely when it has had a bad decade.

That is stricter than “it underperformed recently,” but much less forgiving than “just wait 100 years.”

For example, I would be very willing to say:

Naive small-cap value as popularly understood may be structurally impaired.
Large firms now often have stronger network effects, cheaper financing, better data, better distribution, better AI access, and more ability to acquire or crush smaller competitors. Dimensional itself acknowledges that when the size premium disappoints for a long period, investors naturally search for explanations, including why U.S. small caps have underperformed large caps in recent years. Dimensional

Your proposed regime story is coherent:

Maybe value and small size worked in an industrial, capital-constrained, less scalable economy.
Maybe the next century rewards scale, data, software, brands, platforms, AI infrastructure, and winner-take-most economics.

That is not a crazy story. Stanford’s 2025 AI Index says global corporate AI investment reached $252.3 billion in 2024, with private AI investment up 44.5%, and total AI investment more than thirteenfold higher than in 2014. Stanford HAI A peer-reviewed finance paper also finds that AI-investing firms show higher sales growth, employment growth, market valuations, and product innovation. ScienceDirect That supports your idea that modern technology may structurally favor firms able to deploy scale and intangible capital.

So yes, the next 100 years could be dominated by growth and mega-cap scale.

But there is a catch: once everyone believes that, the price can become the problem.

A factor can fail for two opposite reasons:

1. The underlying economic mechanism disappeared.
   Example: small firms no longer outperform because scale economies, data advantages, regulation, and AI infrastructure favor incumbents.
2. The mechanism still exists, but is overpriced.
   Example: high-quality growth companies are genuinely better businesses, but investors pay so much for them that future returns are poor.

That second point is why I would not automatically conclude “growth wins forever.” A great company is not necessarily a great stock at any price. The market can correctly identify that Microsoft, Nvidia, Amazon, or Alphabet are exceptional businesses, while still overpricing their future cash flows.

This is where factor theory still has some value: not as sacred law, but as a warning against extrapolation.

The statement “large growth companies are economically superior” can be true at the same time as “large growth stocks are bad investments from today’s prices.” Those are different claims.

The harder question is: what evidence would make me abandon the old factors?

For value, I would look for something like this:

Value spreads are not unusually wide, yet value still fails across many definitions, many countries, many sectors, and many rate/inflation environments for several decades.
That would suggest the premium is not merely delayed or crowded, but structurally gone.

For size, I would require less evidence, because size has always been more fragile. I am already more skeptical of standalone “small beats big” than of value, momentum, quality, or profitability. Robeco’s 2024 critique of the Fama-French five-factor model notes several concerns with the model, including issues around factor redundancy and interpretation, which is consistent with the idea that academic factors are not eternal laws. Robeco.com - The investment engineers

For small-cap value specifically, I would say the burden of proof has shifted. It is no longer enough to say “historically it worked.” The defender now needs to explain why small, cheap firms will overcome:

weaker balance sheets, worse access to talent, higher financing costs, less data, weaker distribution, less AI compute, and less regulatory capacity.

That is a much tougher case than in 1975 or 1995.

So I mostly agree with your direction. Where I would push back is on the word “disappeared.” I think it is more precise to say:

Some old factors may have been regime-specific artifacts, not universal premia.

That is very plausible.

And I would say:

The size premium is the most vulnerable to this critique.
Book-value-based value is also vulnerable, because intangible-heavy firms make book equity less meaningful.
Momentum, quality/profitability, and low-risk/defensive effects seem harder to dismiss, because they are tied more to behavior, constraints, leverage, and institutional frictions than to a specific industrial-era balance-sheet structure.

The best analogy to LLMs would be:

Classical factors were like grammar rules: simplified, interpretable, human-designed abstractions.
Modern markets may require high-dimensional, adaptive models because the real structure is too conditional, too nonlinear, and too regime-dependent for five clean factors.

And that is exactly what even traditional quant firms seem to be moving toward. The Financial Times reported in 2025 that AQR has increasingly adopted AI and machine-learning techniques, with Cliff Asness saying the firm had “surrendered more to the machines,” using algorithms to determine factor weightings and detect patterns that may not be fully human-explainable. Financial Times

So my honest position is:

I do not think all factors are dead. But I do think the naive belief in permanent, simple, public factor premia should be dead.

The old view was:

Value and small caps outperform because history and theory say so.

The better view is:

Some characteristics may be rewarded in some regimes, at some valuations, under some constraints, with careful implementation — but no factor deserves unconditional belief.

And your “scale/growth/AI wins” thesis is a serious candidate for the next regime. I would not dismiss it. I would only add: the more widely accepted it becomes, the more future returns depend not on whether it is true, but on whether the truth is already overpaid for.