MIT professor Andrew Lo once made an observation that may be the most precise distinction between finance and physics: “If enough people try to forecast prices, you get no pattern. You get randomness.”
This is not a metaphor. It is a description of a mechanism.
01 / The ball does not change its mind
The stability of physics is about more than rules remaining unchanged.
When you study gravity, gravity does not change because it is being studied. When you use Newton’s laws to trace a projectile, the projectile does not dodge in advance. When you observe the orbital decay of a binary-star system, the orbit does not shift by half a degree because it was observed.
There is no interaction or feedback loop between the researcher and the object of study.
Financial markets are entirely different.
This is not to say that markets are irrational. Quite the opposite: patterns disappear because markets are so rational.
You discover a price pattern and begin exploiting it. Others see your performance and join in. Buying pressure builds, prices move to a new level, and the pattern disappears—not because it was wrong, but because you and your competitors used it up.
In Lo’s terms, randomness is not an assumption about markets. It is the outcome of competition.
If a market is not random enough, that may simply mean too few people are competing in it.
02 / What are you actually betting on?
Once Lo’s paradox is understood, a deeper question emerges: does the object of a bet change because you are betting on it?
The answer depends on what you are modelling.
Fundamental modelling tries to describe the value of an asset itself—its cash flows, industry structure, competitive moat, and management team. These are relatively stable facts. A company’s moat does not change merely because you analyse it.
Expectation modelling is different. It tries to describe other people’s collective judgment about an asset. The object of study is the behaviour of crowds, the formation of consensus, and the path by which information travels through a market.
Lo’s paradox therefore operates most powerfully on expectation modelling.
Fundamentals do not change because you analyse them. Expectations do, because exploiting an expectation is itself a force that changes expectations.
This leads to a crucial distinction: most quantitative strategies are forms of expectation modelling.
Momentum earns from the market underreacting to information. Reversal earns from prices returning after overreaction. Sentiment factors earn from collective irrationality. All exploit patterns in crowd behaviour—and once the crowd realises it is being exploited, its behaviour changes.
That is not a market malfunction. It is the market working as designed.
03 / The dead end in the human response
Lo’s own answer is to control emotion: “The game isn’t predicting patterns. It’s controlling emotions when patterns break.”
That is a human answer. It is reasonable, but insufficient.
The problem is not that emotional control does not matter. It is that human adaptation has a speed limit. The speed at which we regain composure will never match the speed at which a pattern disappears. By the time we recognise that an edge is being arbitraged away, our capital may already be committed to it.
The deeper challenge is that expectation modelling requires us to update our expectation of other people’s expectations in real time. This is not a single layer of recursion. We must infer not only what others think, but also how they interpret our interpretation of what they think. Every additional recursive layer accelerates the consumption of the pattern.
Humans can perform recursive reasoning. They simply cannot perform it fast enough.
The market does not wait.
04 / Evolution must move faster than the market
If Lo’s paradox cannot be avoided—if patterns are inevitably consumed and expectations erase themselves—then the only way forward is not to predict a pattern more accurately. It is to evolve the ability to capture a new pattern before the old one disappears.
That is the significance of code evolution.
Traditional factor mining is search: operators are combined within a predefined space. The search space is closed, everyone explores the same territory, and patterns are consumed more quickly as a result.
Consider an experiment: use the same prompt and operator library, ask GPT through three different API keys to generate 100 factors, and run AST-subtree isomorphism detection on the results. More than 40% of the factor structures are highly similar. This is not a coincidence; it is the inevitable result of a closed search space. AlphaAgent (KDD 2025) independently reported the same phenomenon: factors produced by traditional genetic programming and pure LLM formula search show strikingly high duplication under AST-subtree isomorphism tests.
Code evolution follows a different logic. The system does not merely search inside one space; mutation and selection at the code level allow it to keep moving into new search spaces. It is the difference between finding a route on a map and continuously expanding the map itself.
CogAlpha demonstrates this capability. Its seven-level agent hierarchy—from factor generation and backtest evaluation to code repair—forms an evolutionary loop that can iterate and verify itself. The objective is not to find one better factor, but to build a system that continuously produces better factors.
This is not a faster human. It is a system operating in a different dimension.
At GIM, we have already conducted extensive theoretical work and experimentation in both fundamental modelling and expectation modelling. We will explore that work in future articles. One direction is already clear: when markets process information faster than people collectively can, the system’s adaptation speed must exceed the market’s rate of change.
05 / No Human in the Loop is not a philosophy
When investors hear “No Human in the Loop,” their first reaction is often that it sounds unsafe.
But safety is dynamic.
In a world where a pattern’s half-life is measured in days, requiring human intervention in every decision loop is itself a source of risk. Not because people are not intelligent enough, but because human intervention cannot keep pace with the rate at which patterns are consumed.
If the market pushes back, the act of exploiting a research result becomes the process that destroys it.
The answer therefore cannot simply be better research. The system must be able to run on its own.
This is not because people are unimportant. It is precisely because they are too important: human judgment should be reserved for higher-dimensional responsibilities such as system architecture, risk frameworks, and objective setting—not the last mile of execution in every cycle of pattern decay.
Market reflexivity will not disappear. But a system that can evolve continuously can capture new regularities as reflexivity destroys old ones.
Physical laws do not change their minds. Markets do. That is why you need something that can change faster than the market.
This article reflects GIM’s technical observations and thinking on the asset-management industry. It does not constitute investment advice.