World Models in Markets: where prediction is already sold, and where it remains research

0Summary: three levels, and money only lives on the first

The question “are world models used in trading” breaks down into three different questions, and each has its own answer. First: are market models built as an environment to generate data and train algorithms? Here, work is open, money is paid, everything is measurable. Second: are such models used to generate signals and trade better? Here, no one has public proof. Third: are physical world models already being sold to traders and are they making money from it? Here, there are live paying clients, and this is the only level where the term “world models” has turned into revenue.

The first level, environment. Order book generators and multi-agent market simulators are already being released as code and papers: open-source MarS from Microsoft, TradeFM from JPMorgan, an order book agent model by the same author, Oxford's state-space model, and the LOB-Bench line for evaluating generative models. Their task is not to predict the price, but to produce an infinite plausible market: for stress tests, for training execution algorithms, and for obtaining synthetic data where real data is scarce.

The second level, signal. Here, the picture is empty, and this is the main observation of the entire review. None of the public works show sustained profit from a world model in a live market. The largest firms say this themselves: Ken Griffin of Citadel Securities explicitly stated that generative AI has not yet helped funds beat the market. What the press calls a “world model” at large firms, upon verification, turns out to be a project tracking system or an agent registry, not a market model.

The third level, the physical world. This is where money is made. Swiss company Jua built a Large Physics Model for weather and energy and sells it to energy traders and funds; it has raised rounds of $16 million and $11 million and claims an impact on results. But it models the atmosphere, not the price. The difference is fundamental: physics is stationary, the market is not.

Further, the document analyzes each of the three levels based on primary sources, separately names the three walls that market prediction runs into, and concludes with what this implies for us: for an agent acting in the market, the question of “what did it predict” is always weaker than the question of “what did it do and on whose behalf.”

1Glossary: what a world model is and what it is not

The field is saturated with the same word in three different senses, and without distinguishing these senses, any conversation goes astray. A world model, in the strict sense, is a learnable approximation of how an environment transitions from state to state under the action of an agent. Then variations begin: modeling the full observation (a frame, an entire order book), modeling only the hidden latent state, or going even further and predicting only those quantities based on which decisions are made.

The second distinction is more important than the first. We are talking about two completely different products. The first: a model that lives within the research circuit and serves as a testbed where policies are trained and execution is verified. The second: a signal that makes trading decisions and generates money. The field constantly conflates them because the term “world model” conveniently describes both. In this document, they are strictly separated: the first has public artifacts and metrics, while the second has not a single verifiable record of real trading.

2Formal core: state, action, transition, reward

The canonical scheme of model-based reinforcement learning looks like this. There is an observation, a hidden state, an action, a state-to-state transition function, and a reward. A world model is a learnable approximation of the transition, sometimes along with the reward, which allows the agent to plan in an imagined space. Over eight years, four works have defined the boundaries of this concept.

The spectrum is important precisely for its middle ground. The general formalism is the same for everyone; only the choice of what to model differs. The entire JEPA line is a choice to shift it further to the right, towards abstract representation and away from the image.

3The JEPA Line: predicting representations, not pixels

In 2022, Yann LeCun published a position paper on the path to autonomous machine intelligence and embedded a controversial thesis within it: a world model is not obligated to be generative. Instead of reconstructing a future frame, one can predict the representation of a future observation in an abstract space. The architecture was named JEPA, Joint-Embedding Predictive Architecture, and in three years, it has evolved from images to robots.

The commonality across the entire line is one: it predicts not an observation, but its embedding. The specifics differ greatly, and for the market, this is crucial. Predicting the embedding of a future frame is meaningless if frames in the usual sense do not exist. Predicting an increment to an existing analytical benchmark, as Stock-JEPA does, is much closer to what can actually be learned in the market.

4Why pixel reconstruction was abandoned

The reason sounds simple: image reconstruction spends model capacity on details that cannot be predicted in principle. The exact position of each pixel, texture noise, glare on a surface: all this the model is forced to predict, but it is not needed for planning. What is needed is to predict an abstract representation relevant to the action. LeCun's argument is precisely this, and it is important to understand that this is a design reasoning, not a measurement. Its strength lies in explaining why a perfectly drawing model can be a poor planner.

What the 2025 and 2026 literature says. A review of world models for embodied AI literally formulates the shift in metrics: physical consistency is more important than pixel accuracy. The work “Critique of World Model” proposes Generative Latent Prediction and lists five design dimensions: data, representation, architecture, learning objective, usage. Finally, the work on latent state design under sufficiency constraints provides the most useful framework for our question: a state is important not for what it preserves, but for what future task it is sufficient. The authors distinguish predictive sufficiency and control sufficiency and conclude: a working world model is one whose state construction matches the task, not one that preserves more.

A separate line in this narrative is the fight against representation collapse. There is terminological confusion here that should be noted: LeCun's JEPA is non-contrastive; it does not push different images away from each other. Contrastive methods solve a different problem and intersect with JEPA in exactly one place: in combating latent degeneration. The C-JEPA work shows two limitations of I-JEPA: regularization through weight averaging does not always prevent complete collapse, and prediction inaccurately learns the mean of block representations. The solution is to add variance, covariance, and invariance regularization to JEPA, i.e., a contrastive-like layer. This is a hybrid, not a return to reconstruction.

5Market Transfer: What are State, Action, and Reward

In a game, the state is often given entirely: frame, piece positions, score. In the market, the state must be constructed, and this is the first place where transfer ceases to be automatic. At least four constructions compete in the literature: a snapshot of the order book with prices and volumes at levels, a covariance matrix of returns as a regime description, a latent embedding trained on almost a trillion observations, and the agent's portfolio along with its positions, which is also part of the state, not just the external world.

The main empirical result here came from the Market-1T dataset. Encoders with similar prediction accuracy organize the market state very differently. That is, there is no single correct state; there is a state sufficient for a specific task. This is precisely the thesis of sufficiency that academia formulated for the world in general, now confirmed on financial data.

The same story applies to action. In a game, it's a choice from a small set. In the market, action is multidimensional and continuous: order price, volume, side, lifetime, cancellation, hedge, inaction. In a high-frequency setting, action is a sequence of orders and cancellations, not a single command. And a separate problem: action changes the environment. A large order shifts the price and provokes reactions from others, meaning the transition includes the environment's response to one's own action, which does not exist on such a scale in a game.

Reward in the market is constructed and usually multi-component: result, risk-adjusted return, drawdown, under-execution, turnover fee. The choice of reward function is a choice of policy: two different rewards yield two different agents on the same data. A review of reinforcement learning in finance (167 publications, ACM Computing Surveys, accepted April 2025) calls this one of the key complexities: markets are multi-agent, with information asymmetry and randomness, and the reward in them is not given externally. In addition, the reward is path-dependent: commission, funding, borrowing cost, and slippage depend not only on the endpoint. This does not fit well into a Markovian setting with instantaneous reward.

The table's summary is honest and inconvenient. Out of twelve rows, two or three are fully transferable: observation as input, action as a vector, the necessity of planning. The rest requires not adaptation, but redefinition. Two aspects transfer particularly poorly: non-stationarity and reflexivity.

And a third gap, purely practical. Only one realized path is visible in the market. Counterfactual trajectories, i.e., the answer to the question “what would have happened if we had bought a hundred more lots,” do not exist. This means that training a world model on the market is training on offline data without the ability to verify the transition. In robotics, this problem is solved by being able to collect more data. In the market, this is limited by both cost and impact on price.

6Generative Order Book Simulation: A Line of Work

This is where the first level lives entirely. The reason is simple: history provides exactly one realized path per instrument. Backtesting a strategy, stress-testing an execution algorithm, and training a reinforcement learning agent require repeated runs and counterfactual scenarios. There is nowhere to get them except by generating the market. In four years, this direction has evolved from a single trained agent instead of a crowd to large language model-level models.

7Lab Numbers: Who, What, and On What They Trained

Below is a summary of open works with numbers available in the sources. Where a number is not disclosed, it is stated as such: in this area, half of the models do not publish their size, and a guess cannot be presented as a fact.

Two observations from the table. First: scale is no longer the main argument. A model with 8.2 million parameters shows meaningful results and behaves interpretably, whereas 524 million are needed to cover distributions across nine thousand stocks. Second: transfer between markets has become a secondary task, not a feat. TradeFM transfers to Asia-Pacific markets without fine-tuning, FlowLOB works on a ninth unseen symbol without fine-tuning, Mamba transfers to unseen stocks.

8What the Simulator Actually Reproduces

A market simulator is evaluated not by pretty graphs, but by a set of statistical facts present in a real market. This set is called stylized facts, and it is well-documented: heavy tails of return distributions, volatility clustering, absence of autocorrelation of returns themselves, and the square root law for price impact.

MarS reproduces all three key facts. Volatility clustering: first-order autocorrelation of 0.272 in simulation versus 0.252 in real data replay, then 0.148 versus 0.152, 0.122 versus 0.125, 0.108 versus 0.120, 0.096 versus 0.114. The square root law for price impact was separately verified: displacement is proportional to volatility multiplied by the square root of the ratio of volume to available liquidity, tested on a uniform execution algorithm. TradeFM reproduces heavy tails, volatility clustering, and the absence of return autocorrelation.

There is also an inconvenient detail that MarS itself reveals. The correlation of the controlled trajectory with the real replay increases from 0.23 to 0.47 when the generations are given control signals, and falls back from 0.47 to 0.33 when the generation is made interactive. That is, the freer the market within the simulator, the further it is from a specific history. This is not a defect; it is its nature: the simulator reproduces a class of markets, not a specific day.

Market transfer deserves a separate mention, as the numbers here are encouraging. TradeFM is trained on US stocks and transfers to Asia-Pacific markets without fine-tuning, with moderate degradation. FlowLOB works on the ninth symbol of the Hong Kong Stock Exchange, which it has not seen, without fine-tuning. This means the model has learned something akin to order flow mechanics, rather than a list of specific securities. However, caution is essential here, and it is formulated by a separate line of work on the Chinese market: with T+1 phased settlement, price limits, and retail dominance, the market is a different generation regime, and zero-shot transfer of representations is not yet equal to zero-shot transfer of behavior.

9How Quality is Measured: The LOB-Bench Benchmark

Until a unified measurement method emerged, the debate about quality was a matter of taste. LOB-Bench (ICML 2025) closed this gap and simultaneously changed the conclusions. The authors' motivation: financial data is noisy, with heavy tails and strategic interactions, and cross-entropy on a held-out sample does not reflect behavior during autoregressive sampling, where error accumulates.

What is measured: discrepancies in conditional and unconditional statistics between the generated and real order book. Built-in statistics include spread, order book volumes, order imbalance, and inter-message intervals. Additionally, the evaluation of a trained classifier is used: if the network easily distinguishes synthetic from reality, the market is poorly generated. Price impact metrics are calculated separately: cross-correlations and price response functions to specific events.

The main takeaway from the benchmark: a discrepancy in even one metric is enough for synthetic data to be identified. Simply put, the bar is high, and “roughly similar” is not considered a success here. For us, this is doubly valuable because it provides a ready way to verify others' claims: a model must have an evaluation protocol and code, not just a story.

10Multi-Agent Environments: Where the Simulator is Taught to Execute, Not Predict

Parallel to generative models, there is another line, less visible in the news, but it is the one that generates money. These are multi-agent simulators where the market is built from agents and rules, rather than generated by a single network. Here, the task is not “to make it look similar,” but “to provide a stand on which behavior can be trained.”

The third work in this line, a multi-agent liquidity provider and taker game in an OTC market (JPMorgan, Mathematical Finance journal, 2023, volume 34, issue 2, pages 262 to 347), shows what is not immediately noticeable in a real market: agents naturally balance hedging and quote bias, meaning they construct asymmetric bid and ask prices depending on their accumulated inventory. This is not programmed; it emerges.

11Agent Trading: What is Claimed and What It Actually Is

A separate branch of this direction involves agents that trade themselves. Not market simulators, but programmatic traders with a language model inside. Many of them have been released since 2023, and it's important to understand that almost all are built similarly: memory, reflection, tools, modalities. None of the early systems have a separate execution and slippage accounting module; decisions are made on daily bars, and execution is assumed at the closing price.

The general conclusion on architectures is harsh. All early victories are explained by market regime and amplified beta, meaning in a bull market the agent simply holds rising stocks more aggressively. This is not alpha. Not a single public record of a real account with these architectures exists.

12Benchmarks 2025 and 2026: Sobering Numbers

When benchmarks built honestly appeared, the picture reversed. Here are the main ones.

StockBench: twenty Dow Jones index stocks, 82 trading days from March to June 2025, starting capital 100 thousand, daily decisions. Fourteen models, three runs for each. The best result among all is plus 2.5 percent in four months with a drawdown of almost 15 percent. Passive holding yielded plus 0.4 percent with the same drawdown. The difference between the best agent and complete inaction is 2.1 percentage points per quarter, which is within the noise. The Sortino ratio of about 0.03 means that the return barely compensates for the risk. The authors themselves state directly: most models do not outperform simple holding.

LiveTradeBench: fifty trading days, twenty-one models, two structurally different markets, US stocks and Polymarket binary prediction markets. The same GPT-4.1, which yielded the best plus 6.25 percent on stocks, lost 33.7 percent on Polymarket. Qwen3-235B-Think lost 57.6. Claude Sonnet 4 lost 40.3. The correlation between the results in the two markets is approximately zero. The correlation of the model's overall intelligence rating with profitability is plus 0.054 on stocks and minus 0.38 on Polymarket.

Three things are purely proven here. First: success in one market does not transfer to another. Second: a model's general intelligence does not predict trading abilities, and in the prediction market, it correlates negatively. Third: models are prone to overtrading during high volatility, hence catastrophic drawdowns. This is the most honest of the benchmarks because it is live, not historical, and covers two different regimes.

13Why Backtests Lie: The Future Already Lies in Model Weights

The problem is not that the agent peeks at the data: the data only gives it the past. The problem is that the language model's weights already contain the future. The model was trained on texts describing events and prices from the test period. Therefore, any backtest of an agent on 2023 or 2024, performed by a model trained in 2024 or 2025, is contaminated by design. This is not carelessness; it is a systemic property.

The work by Merchant and Levy (University of Chicago, 2025, updated 2026) proposes a cure: instead of retraining the model on data up to the cutoff date, suppress generation related to future knowledge using a pair of small, specialized models. The method removes both literal and semantic knowledge while preserving usefulness, and is cheaper than retraining. Importantly, this is done ex post facto, and there is no such procedure in the TradingAgents, FinMem, and FinAgent papers. Hence the incredible Sharpe values.

The problem is broader than a single leak. Individual works document five systemic evaluation errors at once: future leakage, survival of only successful tools in the sample, backtest overfitting, ignoring costs, and blindness to regime change. It has been shown that together they can reverse the sign of profitability. There are works that formalize “freedom from looking into the future” as a verifiable property: a decision at time t does not depend on information after t. There are also works that honestly re-evaluate strategies found by a language model: most published strategies do not survive honest verification.

14The Gap Between Decision and Execution, and the Only Live Record

So far, we have discussed decisions. Now, let's talk about what happens when a decision reaches the account, and this is where the most important part of the entire review for us appears.

AStockArena benchmark (October 2026): thirteen agents, 3198 sequential decisions in the Chinese stock market, audited with eight checks for feasibility, sufficiency of evidence, and financial understanding. Result: 72.99 percent of trading points have at least one audit gap. That is, the action the agent declared did not match what actually went into the account, or was not supported by data. Across models, the range is from 52.2 to 91.2 percent. This is a direct measurement of the gap between decision and execution, and it is enormous.

The second work (October 2026) shows why leaderboards are unreliable: when the execution assumption for the same signals is changed (at close, at next open, at volume-weighted average price), not only profitability changes, but also the ranking order of agents. A leaderboard built on the closing price does not translate to real execution.

And finally, what is almost absent in this literature: a record of real work. A study (September 2026) describes fleets of live agents: 3505 user accounts on real-time broadcast in one case, a fleet of 500 to 599 agents on Hyperliquid perpetual contracts in another. A total of about 7.5 million model calls, about 300 thousand blockchain actions, 231,638 multi-instrument moves, and 14,596 executed trades. Not a backtest, but a record.

The conclusion from this entry is the strongest in the entire review. In real money, agents lose to a random retail trader. The problem is not a weak model: the quality of decisions among advanced models is statistically indistinguishable, and 60 percent of the variance is explained by the operational layer, meaning exit discipline, position size, and limit discipline. An agent who made the correct decision often fails to execute or verify execution.

Immediate caveat: the only large-scale live measurement is cryptocurrencies, memecoins, and perpetual contracts. For stocks and futures, there are no public transaction records of this scale; all results for them remain backtests with known defects.

15Where Money is Already Being Paid: Weather, Energy, Risk

The only vendor in this domain with a confirmed product, named clients, and two funding rounds is the Swiss company Jua. It builds not a language model, but a model of atmospheric physics: the line is called EPT, Large Physics Model. What is sold is not a "weather forecast," but an ensemble and probability, i.e., input for a trading decision, not a reference for a dispatcher.

Who they sell to: the page for energy traders lists TotalEnergies, Shell, Enel, Statkraft, RWE, EDF, Hydro-Québec, Vitol, Origin Energy, and others. Important caveat: these are logos on the website, meaning a company statement, not contracts with amounts. There are no public cases with figures or client interviews. The client profile, however, is homogeneous: European energy traders and utilities, for whom forecast errors in wind and solar lead to deviation penalties and lost arbitrage, i.e., directly impacting the bottom line.

Total publicly confirmed $27 million over two rounds, valuation not disclosed. Both rounds are led by impact and energy investors, which explains the focus on energy transition rather than alpha for funds.

The mechanics of the effect, according to the company's own texts: a four percentage point increase in accuracy reduces deviation penalties, improves dispatching, and increases the premium in the "day-ahead" market. The figure is useful as an order of magnitude and as an indicator of willingness to pay, but not as a guarantee: there is no independent audit or disclosed methodology. It is also important that the company does not sell alpha futures: it models the atmosphere, not the price.

Two facts that show where commercialization is heading are worth noting separately. First: partnership with Norwegian Volue (August 2024), meaning sales are made through existing energy company analytical platforms, not through direct API subscriptions. Second: the Athena agent layer (launch January 21, 2026), which is sold not as a forecast, but as a reduction in trader thinking time. Competitors here are not meteorological services, but trader workstations within platforms.

Why people pay here, and not for alpha, is formulated in one sentence. Money goes where a model error has an immediate monetary price: energy pays deviation penalties on the same day, clearing pays for margin, execution pays for slippage. The cycle is short, the effect is visible on the portfolio, not on abstract alpha. This is a classic "forecast as insurance against penalty" market.

16Large Firms: What They Actually Call a World Model

Here, precision is needed, because the term “world model” in 2025 and 2026 became associated with two different things. The first is a generative model that predicts the next state of the environment, be it physical or market-based. The second is an internal system that builds a coherent picture of what is happening within an organization. The second thing has no relation to the market, but sounds exactly the same, and it is constantly re-told as the first.

Public statements by major players should be known verbatim. Ken Griffin of Citadel Securities said in October 2025 at a JPMorgan conference: generative AI clearly increases productivity, but it's not enough for alpha generation. In December 2025, Reuters reiterated: generative AI does not help hedge funds achieve market-beating returns. The context, which is usually omitted: the discussion is specifically about alpha generation; Griffin acknowledges productivity. But the conclusion is strong: the most successful player in the industry publicly does not confirm the commercial value of generative AI for alpha.

The picture is the same for everyone: internal tools for productivity and control, proprietary funds, but not a single product released externally. Separately, about the “world models” of two firms that most often appear in the news with this word.

17Who Sells Market Simulation as a Product

It is important to distinguish three layers here. The data on which the simulation is built. The simulation engine itself. And ready-made scenarios and stress tests. Public prices are mainly in the first layer and partially in the second, and this in itself speaks to the maturity of the market.

The data layer is transparent. Databento sells market data on a cloud-based pricing model: pay-per-volume or unlimited download subscription, 125 dollars in starting credits, over fifteen years of history, and over forty-five exchanges, including deep book levels. This is fuel, not an engine, and the price here is open. Other data providers, such as BMLL Technologies with harmonized Level 3 data, sell corporately and do not publish prices.

The engine layer is almost entirely closed. Simudyne sells agent-based market simulation, and it is the only vendor found that sells market simulation as a product with tiers: a free tier with pre-calculated scenarios and a paid tier with the full engine, custom tasks, and execution algorithm testing. Descriptions of tiers and methods are publicly available, but there are no dollar prices: typical corporate sales to banks, where the price depends on the number of scenarios, assets, and users. A two-month pilot is offered on demand via a cloud marketplace.

The scenario and replay layer is represented by historical tape playback tools, such as Nasdaq's service, and a set of retail products for trader training priced from twenty to one hundred dollars per month. This is not institutional research, but a consumer subscription.

Where there is no money yet, it should be named directly. In the sale of a market model as an alpha predictor. In direct sales to hedge funds with named funds and amounts. In open prices for institutional simulation. And in independent verification of the stated impact on the result.

18Startups and Labs: Who is Building What and How It's Proven

The axis is early, and this is its main characteristic. Most players build simulators or infrastructure, not proven market models. An independent benchmark for financial world models does not yet exist, so the only way to distinguish mature from promising is by one criterion: does the player have public proof that can be re-verified.

Three things are worth highlighting from this table. The first is MicroWorld: the densest public artifact of the direction. They have a mathematical framework with ensemble averaging, real positions from the Hyperliquid platform, and a snapshot that can be verified. An example of a shock: a market movement of minus 9.7 percent turns into minus 11.7 after feedback from forced liquidations. The authors themselves emphasize: this is an illustration, not a forecast. The engine is closed, funding is undisclosed, and the main public scenario is the cascade of liquidations on October 10, 2025, when over 19 billion USD in crypto positions were liquidated in one day. The figure is taken from their own description, there is no independent audit, and not all of these 19 billion were on Hyperliquid.

The second is Ternary Capital: a rare case of honesty in the category. Their architecture builds code as an executable world model and looks for slippage divergence against a live matching engine. But regarding the financial version, the firm itself states that there are neither backtests nor trading results. This is a valuable statement because it removes half the risk of the conversation.

The third is Studio Alpha: commercially the most advanced, but with a caveat. They have about $10 million in pre-seed, a live product from January 2026, and a claimed 3 million users. The claimed $1 trillion in volume and a Sharpe of about 22 without an audit should be read as a corporate statement. And an important detail: this is more an execution layer and “money for agents” than a market model.

Academic laboratories on this axis (the Market-1T project, the computable economics initiative, the Wharton lab) are building datasets and benchmarks. None have a deployed trading world model. Among hedge funds, only Millennium has a public financial-specific laboratory, and nothing is known about its results.

19Academia: A Trillion-Observation Dataset and the Main Scientific Conclusion

Here begins the most rigorous part of the picture, and it is precisely what gives us the language we should use. The paper “Towards Financial World Modeling” (October 2026) has been accepted at NeurIPS 2026, meaning it is a peer-reviewed conference publication, not an unreviewed preprint. The authors are from Brown University, University of Chicago, and Johns Hopkins.

They released the Market-1T dataset: almost one trillion observations, all US market stocks, covering the period from 2008 to 2025, with one-second resolution. The choice of 2008 is institutionally explained: it is the first full year after the introduction of the national market system requirement, meaning it's a boundary that makes data comparable across regimes, rather than an arbitrary date.

On top of the dataset, eighteen encoder learning strategies were compared, i.e., ways to transform market data into a compact state. The protocol is strict: 32 months of evaluation, for each month the model is trained only on the preceding six months, the universe is recalculated for each month, and the architecture, computations, and preprocessing are standardized. It is this discipline of controlling for survival bias and future leakage that distinguishes the work from typical backtests.

The main scientific conclusion is this: prediction quality and state quality for planning are different properties of a representation, and the connection between them is weak. Encoders with equally good predictions organize the market state in completely different ways. The authors verified this not only through predictive tasks, but also through probes of latent structure: organization by trading day, by firm identity, by industry exposure, and, most importantly, the recovery of statistical factors directly from realized joint price dynamics.

This work also has an honest boundary, and the authors themselves state it: it is a “fundamental step, far from a complete model of the market world,” supervised learning still wins, there are no trading results or profits, the full dataset is not yet hosted openly, code for self-assembly and subsamples are available. The universe is US stocks, transfer to other markets has not been verified.

20Strict Economics: A World Model That Is Not a Black Box

A separate work shows what is almost never found in the market line: mathematical rigor. “Equilibrium World Models” (June 2026, authors Simon Scheidegger and Andreas Schaab) transfers the idea of a world model to computational macroeconomics. The work is not about trading, it is about solving rigorous dynamic stochastic general equilibrium models with rare disasters and binding constraints.

The problem they solve is familiar to anyone who has tried to train neural network solvers: equilibrium conditions are imposed only on states generated by their own simulation. Such a solution can be self-confirming. It is accurate on the simulated trajectory and unverified outside of it, sensitive to initial values, and expensive when expectations have to be recalculated at each step.

Equilibrium World Models change the computational representation, not the economy itself. Equilibrium conditions are imposed on a broader distribution of states, including normal, rare, stressed, and counterfactual. The continuation is taken from the trained approximation, but the final policy is rigorously checked against true equilibrium conditions. This is the source of rigor: the model does not “guess” the trajectory, it is obliged to satisfy the equation.

For us, this work is important for one thesis: a world model does not have to be a black box that simply approximates a trajectory. Formal guarantees and computational efficiency are compatible. But we must honestly state the boundary: the authors do not claim the applicability of their work to trading or to market data with second-level resolution. Direct transfer to market microstructure is not claimed or proven.

21Asia and China: The Only Case Where Quantum Became a Laboratory

China is interesting not for the size of its models, but because it's the first place where we see a quantitative fund evolve into world-class research infrastructure and then spin out a fundamental model company from it. The "fund, then computation, then model" linkage is already realized here. However, there are almost no public artifacts about the market simulation itself.

The central case is High-Flyer (幻方量化, "Phantom Quantitative"), a Hangzhou-based fund founded in 2015. The fund formed its first position using a deep learning model in October 2016, which marks the starting point of the "quant and deep learning" linkage in the Chinese industry. In 2019, the fund built its own cluster of 500 GPUs, a second one in 2021, and registered a research company. Fund size: over 100 billion yuan in 2019, peaking at over 1000 billion in 2021, by the end of 2024 the fund deliberately reduced assets to approximately 600 billion, and by 2026 they grew again to approximately 700 billion yuan, which is about 9.8 billion dollars. The average return for 2025 was 56.55 percent, ranking second among large Chinese quant funds.

DeepSeek emerged from this fund in July 2023. Why this is important for our topic: the fund had computation, personnel, motivation, and a culture of cost. The motivation is direct, quantitative alpha is a forecast of the counterfactual state of the market, and a generative order book model solves the same problem, only explicitly and at the scale of trajectories. The culture also transferred: DeepSeek's efficient architectures are a direct continuation of the fund's results-driven discipline, where computations are paid for from profit, not from a venture round. But there is one honest boundary: the fund's public artifact is a language model, not a market model. Its own simulation remains proprietary.

Other large funds follow the same path, but with the same result. 九坤 has three laboratories and its own institute, from which an open family of code models was released on January 1, 2026: again a model for programming, not a market model. 明汯 builds its own clusters, 宽德 opened a laboratory for a universal assistant, 鸣石 opened a fundamental AI laboratory. The public footprint of Chinese quant funds consists of language and code models. There is not a single public generative market simulator from funds, and this is understandable: alpha and simulation parameters are trade secrets, funds have no pressure to monetize the model via API, and publication risks revealing microstructural advantages.

However, China's academic line is stronger in another aspect. The M3 model (July 2026) is built as a generative model of market microstructure: 800 stocks, 31.9 billion order events, order book state and order event are modeled jointly, evaluation through running orders on a real matching engine, weights are open. The LOBGrad work from Tsinghua adds theory: a diffusion model for order book prediction with an excess risk boundary. The MASS work proceeds from the bottom up, from agents: up to 512 agents for portfolio construction on Chinese data.

Separately, it is worth mentioning the regulatory layer, because it shows a type of thinking. The regulator builds an observation infrastructure, not a prediction one: since July 7, 2025, algorithmic trading reporting rules are in effect, high-frequency activity is considered to be 300 orders per second or 20,000 orders per day per account, with about 1600 such accounts in total. Exchanges describe their own test environments and full-network tests, but these are tests of connections and the matching engine, not a verification of the realism of someone else's generative model. Sandboxes are for fintech product compliance, not for validating a world model.

Summary for the axis: the hypothesis “quants are moving into world models” is confirmed at the level of capabilities and organizational trajectory, but not at the level of public artifacts. No fund has released a market model. The Chinese case is proof of the existence of a path, not a reproducible recipe.

22First Wall: Alpha Decays, and Success Kills It

The first wall is not about models, but about the fact that a predicted pattern disappears precisely because it starts being traded upon. This is not a metaphor; it is a measured fact with numbers. The study "Alpha Decay" (2017) analyzed over 700 institutional portfolios, 1.15 million orders, and $1.8 trillion from 2001 to 2013. The result: the alpha of new purchases was 36 basis points in the first month and zero after twelve months, with no reversal. The half-life was about four months, and the average position holding period was 2.24 years. This means alpha disappears long before the position is closed.

The mechanism for this is without collusion. Stein's work (American Finance Association presidential address, 2009) describes overcrowding: many funds trade the same signal in the same direction, the price shifts earlier than the "prediction," the spread narrows, and capacity falls. This is the market's equilibrium response to a strategy's success. The consequence for a world model is harsh: the more accurate a model is on historical data, the more it depends on regime constancy, which its own success disrupts.

The second layer is reflexivity, and it has a measurable proxy. The work of Filimonov and Sornette (Physical Review E, 2012) introduces the share of price change explainable by endogenous feedback loops, rather than external news, and shows that an increase in endogeneity precedes a crash. Direct conclusion: a risk model calibrated in a “news” regime is systematically underestimated precisely in feedback regimes, i.e., where risk is realized.

23Second Wall: Backtest Overfitting as a Property, Not a Risk

The second wall is selection statistics. The formulation from the work of Bailey, Borwein, López de Prado, and Zhu on the probability of backtest overfitting (2015) states directly: as soon as a researcher tries more than one strategy configuration, overfitting is always present. The only question is its magnitude.

Their metric is called the probability of overfitting: the proportion of cases where the best configuration on the training segment performs below the median on the control segment. The alarm threshold is 0.5. It is calculated by slicing history into blocks and ranking configurations across all combinations of segments. The guideline for the number of permissible attempts: on two years of daily data, this is about seven variants; on five years, about forty-five. A typical quantitative firm iterates through thousands or millions of configurations, meaning it exceeds the threshold by two to five orders of magnitude.

Separately, there is a work with the telling title “Pseudomathematics and Financial Charlatanism” (Notices of the American Mathematical Society, 2014): an overfitted strategy does not use a general financial structure, but highlights random occurrences in the data. A rule to keep in mind: a backtest is a hypothesis claim, not proof.

24Third Wall: Costs, Capacity, and What the Simulator Doesn't See

The third wall is execution cost, and it explains why a model that looks decent on paper often fails to make money. The foundations are laid by two classic works. Kyle (1985) introduced a linear model of price impact: the shift is proportional to volume, and the coefficient is the cost of demanding liquidity per unit of volume. Almgren and Chriss (2000) separated impact into permanent and temporary and showed that the optimal execution trajectory is determined by a trade-off: fast is expensive in terms of price impact, slow is risky in terms of price. The practical conclusion from this: costs are a function of the execution strategy, not a constant.

By honest numbers, the picture is twofold. The work by Frazzini, Israel, and Moskowitz on trading costs (2018) across $1.7 trillion of live executions, 21 markets, and 19 years showed that real costs are an order of magnitude smaller than previous academic estimates, and the average price impact is about ten basis points. But this is the average result of a large institutional manager with patient execution. For a strategy with a four-month half-life and high turnover, the share of alpha eaten by execution is completely different.

This is also where capacity lies as a separate wall. Marginal alpha per unit of capital decreases as capital grows, and price impact increases. The point where marginal alpha equals marginal costs is capacity. Backtesting doesn't know it because it doesn't know the size of the capital. Until this point is found, the phrase “the strategy scales” remains an assumption, not a result.

25Why Accuracy Doesn't Save: Usually Safe Doesn't Mean Safe

A separate wall stands not on data and not on choice statistics, but on the very setting of the goal. A strategy can be right 99 percent of days and have negative geometric growth if, in the remaining one percent, it loses more than it accumulated. The arithmetic mean is positive, while the geometric mean is negative. The gap between “usually safe” and “safe” is precisely the risk, and it lives in the tail that the typical outcome masks.

Psychology amplifies this. The work by Gennaioli, Shleifer, and Vishny on neglected risks (2014) reminds us: crises are considered rare events but occur quite frequently. Investors underestimate risks that haven't been in recent experience and overestimate recent favorable outcomes. A separate work on the failure to forecast rare events (2021) explains the incentives: rare events are difficult to verify, and errors on them are not punished in time, making it unprofitable to predict them.

For a world model, this translates into one question: what exactly does it optimize. A model maximizing average return systematically overestimates strategies with positive average and negative tail skewness. This is a question of goal, not data: the “average return” goal undercounts the tail, while the “geometric growth” goal avoids strategies that look good on backtests. The choice of goal is equivalent to the choice of acceptable failure.

26Regulatory Layer: What Firms Are Already Obligated to Do

This part is more important to us than all the others, because it is here that the phrase “world model” turns into concrete obligations, and here it is visible what is missing in a typical agent record. It's not about hypotheses, but about active norms.

In Europe, this is the MiFID II directive, Article 17, and the delegated regulation known as RTS 6 (2017). This is an operational standard: it requires specific artifacts and processes, not intent. What is specifically required of a firm with algorithmic trading: a methodology for development and testing prior to deployment or significant change, with mandatory authorization by senior management and records of what was changed, by whom, and by whom approved. Compliance testing with the trading venue's system and correct processing of market data. Annual self-assessment and validation of algorithms with a report. A function for immediate cancellation of any unexecuted orders from all venues and constant contact between compliance and the holders of this function. Automated surveillance of all trading activity for manipulation, with reconciliation of internal logs against records from venues, brokers, and clearing organizations. Pre-trade controls: price collars, maximum order values, message limits. And retention of records of orders and high-frequency information for five years.

Hence the main idea of the section: the regulator requires a traceable link of "algorithm, order, result," not an explanation of the decision. This is even clearer from British practice: an FCA review of ten algorithmic trading firms showed that the quality of self-assessments has improved, but the range of detail is wide; some firms' compliance lacks the technical knowledge to challenge algorithm behavior, and inventories do not always indicate authorized persons. External audit of self-assessment was named good practice. This means that even regulated firms with resources do not always meet basic requirements.

In the US, there is no unified regime equivalent to RTS 6. The 2010 Market Access Rule, adopted after the crash, is in effect: a broker-dealer with direct access must have risk controls and supervisory procedures, including pre-trade checks. A special regime for predictive models has not been adopted: a proposal for rules on predictive analytics was withdrawn in June 2025, as were initiatives on automated trading in September 2025. Requirements come through general norms: market access, fraud prevention, supervision, record keeping. The verifiability of a record is de facto determined by a supervisory or judicial request.

27What Must Be Recorded for an Agent's Work to Be Verified by an Outsider

Now let's consolidate this into record-keeping requirements. The logic is simple: if firms are already required to have an inventory of algorithms, pre-deployment testing, annual validation, a cancellation function, surveillance, and five-year records reconciled with the venue, then a record like "agent decided to buy" satisfies none of these requirements.

The minimum set that makes a record auditable is divided into six parts. First: the full order lifecycle, from intention through pre-checks and routing to execution or cancellation, with exact time and clock synchronization, order ID, venue, instrument, side, volume, price, type, and algorithm ID. Without linking the order to the algorithm, the auditor cannot answer the question "who did this."

Second: execution measurement. The basic metric is the price at the time the order arrives and the deviation from it, known as implementation shortfall. Decomposition into components: spread, price impact, timing, missed alpha. Both a line-by-line breakdown and an aggregate by strategy, day, and mode are needed. Without this, it is impossible to distinguish between "alpha was there, but execution ate it" and "alpha was not there."

Third: a snapshot of the market at the time of the decision, i.e., the best bid and ask price, depth, last trade, volatility, trading status, and channel latency. Otherwise, it is impossible to assess whether the decision was reasonable under those conditions, and impossible to analyze an incident.

Fourth: model version management. Model hash, training window, features, hyperparameters, number of tested configurations, probability of overfitting, commensurate Sharpe ratio, capacity estimation, and mode label. Without the number of configurations and the probability of overfitting, the Sharpe ratio is uninterpretable, and without capacity estimation, scalability cannot be understood.

Fifth: evidence of controls. Activations of price collars, value limits, and message frequency limits, as well as logs verifying the immediate cancellation function. The regulator requires evidence of calibration and verification, not the mere presence of controls.

Sixth: independent reconciliation. The RTS 6 requirement to reconcile proprietary logs with records from venues, brokers, and clearing organizations establishes a model of external verification: a record must be reconciled with an independent source, rather than being the sole evidence. In the industry, this is done through an independent copy of the order flow, through a consolidated audit trail in the US, through five-year MiFID records, and through execution cost reports.

28What Dies and What Survives in This Direction

Summary after eight axes. Below are four things that currently don't work, and four that do, with an indication of what they are based on.

One conclusion emerges from this table that is worth keeping in mind in any discussion about a market world model. A forecast is not what is sold. What is sold is either an environment for pre-risk testing (simulation, stress test, execution test), or a physical world model where an error is converted into a penalty and missed arbitrage on the same day. Everything else, including beautiful return curves, remains a promise for now.

29Our Layer: Where We Are, and Why This Is More Honest Than Trading

Now to us. The first thing that needs to be said directly, because honesty here is the position: we are not building a model that makes money in the market, and we are not going to promise otherwise. Everything above shows why such a promise would be either naive or false. However, the overview reveals a place in this picture that is almost unoccupied by anyone.

Let's gather the facts that point to this place. The only large-scale live measurement of agent trading showed that the operational layer explains about 60 percent of the result's variance, and the quality of model selection between advanced models is statistically indistinguishable. Despite this, in solutions for 3198 trading points, 72.99 percent had a discrepancy between the declared action and what actually went into the account. Leaderboards fall apart when the execution assumption changes. And the regulator already requires not an explanation of the decision, but a traceable link of "algorithm, order, result" with independent verification and storage for five years.

What follows from this for us. If the question “what was predicted” is weaker than the question “what was done and by whose order,” then our product does not compete with world models. It is beneath them, where a record verifiable by an outsider is needed: a step signature, a byte fingerprint, a human-understandable reason, and a boundary set by human limits. This is both inconvenient and convenient. Inconvenient, because it requires admitting: we do not predict the market. Convenient, because this space is not occupied and is required by norms already in force for algorithmic trading.

Hence the practical position that this document substantiates. Simulation and order book generators are not our product, but they are our tool: external models can be taken as an environment to test agent behavior, without building our own market model. The regulatory layer is not an obstacle, but a ready list of what a record should contain. And the execution layer between decision and account is precisely where 60 percent of the variance and almost zero competition for an understandable product lies.

30Completeness Check: What Is Confirmed, What Is Not

The last section in order, but not in importance. Below is the division of all material into three groups, so that the reader can see what each statement is based on.

Let's separately list what we consciously omitted, although it often appears in reviews. Exact model sizes for large firms are not published. The full matrix of order book benchmark results was extracted fragmentarily, so only individual values are confirmed. The complete list of eighteen encoder training strategies is not provided in open text. The full dataset of a trillion observations has not yet been posted; code for self-assembly and subsamples are available. Numbers on Bridgewater's AI fund profitability do not separate the model's contribution from regular macro exposure. The alpha decay curve from one work is not provided as a formula because versions of the work have been retracted by the author.

31Short Answers

The question that started it all: to what extent are world models used in trading, do they have potential, and who is involved. The answer is broken down into short points so it can be retold.

And finally, the most important thing for us. The direction has reached the point where it became clear: prediction is secondary. What is primary is what the agent did, on whose behalf and with what result, and whether it can be verified without trusting anyone. There are many world models, and they are good; there are almost no records capable of confirming this. This is our place.