If LLMs Know About
Reflexivity, Do They
Forecast Better?

Almost every paper on LLMs as financial forecasters treats the market as an exogenous system. The model looks at data, makes a prediction, and the prediction has no effect on what it's predicting. This is the standard assumption. It is also, in a narrow but important sense, false.

George Soros spent decades arguing the opposite. His theory of reflexivity holds that market prices don't just reflect fundamentals — they feed back into them. Investor beliefs shape prices, prices shape corporate behaviour, corporate behaviour shapes fundamentals, fundamentals shape beliefs. The loop is closed. A forecaster is not an outside observer. They are a participant in the system they are trying to predict.

A paper published in May 2026 asked a simple but underexplored question: if you tell a frontier LLM about reflexivity — explicitly, as context — does it forecast boom-bust cycles better?

// The experimental design

The setup is careful in ways that matter. Historical price and market data from two episodes — the dot-com bubble and the GFC — are anonymized and normalized before being fed to the models. The normalization is intended to prevent the models from recognizing the specific historical period and pattern-matching on memorized outcomes. Whether it succeeds is one of the central questions I'll return to.

Each model is evaluated under four conditions of increasing theoretical scaffolding:

Condition 1 (baseline) — raw normalized price data, no additional context. The model forecasts direction from the data alone.

Condition 2 — the model is told it is making a financial forecast during a market cycle.

Condition 3 — the model receives a description of reflexivity theory: that prices and fundamentals form a feedback loop, and that self-reinforcing dynamics cause trends to persist longer than fundamental value would suggest.

Condition 4 — the model receives the full Soros reflexivity framework including boom-bust cycle structure: the initial trend, its amplification through feedback, the moment of inflection, and the eventual correction.

// What happens when you add the theory

Directional accuracy improves as the scaffolding accumulates. The models with full reflexivity context (C4) outperform the baseline (C1) on both episodes. The improvement is not marginal — it's consistent across all three models and both historical periods. The Sharpe ratio of the implied long/cash strategy also increases, which matters: accuracy that doesn't survive trading costs is the problem the Zhang & Zhang paper documented extensively, and here the economic signal is positive even at C4.

The dot-com bubble shows stronger improvement than the GFC. This is plausible: the dot-com bubble is a cleaner example of Soros's original framework — a clear self-reinforcing narrative (internet stocks must rise because the internet changes everything), amplified by retail participation, followed by a sharp inflection. The GFC is a more complex multi-stage event with regulatory, liquidity, and contagion dynamics that reflexivity theory doesn't fully capture. A model given the reflexivity frame may apply it more confidently to the cleaner case.

// The confound the paper can't fully resolve

Here's the problem. Anonymization and normalization guard against the most obvious form of memorization — the model recognizing "1999 NASDAQ" from the numbers and recalling the outcome. But they don't guard against a subtler form: the model recognizing the shape of a boom-bust cycle from normalized data and retrieving learned associations between that shape and its typical resolution.

LLMs have read a great deal of financial history. The dot-com bubble has a recognizable pattern — exponential growth phase, plateau, rapid decline — that is structurally similar across many historical episodes. When the paper provides the Soros framework in C4, it is simultaneously providing two things: a theory for why prices behave as they do, and a narrative template that may help the model recognize which historical episodes the anonymized data most resembles.

The paper's authors acknowledge the memorization concern and address it through normalization. That's the right move. But the narrative-recognition confound is harder — it doesn't require the model to remember specific numbers, only to recognize structural patterns. And LLMs trained on internet text have seen boom-bust cycle narratives described hundreds of thousands of times, across every bubble in recorded financial history.

// Three models, one pattern

One of the more interesting findings is that the improvement holds across all three frontier models — GPT-5, Claude Sonnet 4.6, and Gemini 3 Pro — despite their different architectures and training regimes. The direction of improvement is consistent even if the magnitude differs. This is some evidence against a pure memorization explanation: you'd expect different models to memorize differently, and if the effect were purely about recognizing historical patterns, you'd expect more variance between models with different training data compositions.

It's not conclusive — all three models have been trained on significant overlapping financial text — but it does suggest that something more general than a specific memorized association is driving the result. Whether that general thing is "genuine understanding of reflexivity" or "general competence at recognizing boom-bust narrative structure" remains open.

// What this is actually evidence for

Setting the confound aside, the paper establishes something useful. Providing theoretical context to LLMs in a financial forecasting task improves performance. This matters for practitioners thinking about prompt design, not just researchers thinking about market theory.

The implication isn't "give your trading model Soros." It's narrower and more actionable: when asking an LLM to forecast a dynamic with a known structural mechanism — a feedback loop, a mean-reversion process, a regime shift pattern — providing the theoretical description of that mechanism improves the output. The model has already read every paper on reflexivity, every paper on momentum, every paper on mean reversion. The right context prompt may unlock that knowledge in a more targeted way than asking it to reason from data alone.

Whether the mechanism is genuine theoretical reasoning or sophisticated pattern priming probably doesn't matter for deployment. What matters is whether it generalises to novel episodes. That experiment hasn't been run yet.