ProjectAtlas
Back to the presentation

Technology and Economics

Can World Models Improve Financial Forecasting?

Yes, progressively, and the horizon is widening. Research published through 2026 shows generative market simulators reproducing order-book behaviour and stylised market facts, financial foundation models forecasting across instruments from one pretrained model, and deep order-flow models extracting signal over multiple horizons. A World Model adds what a standalone predictor lacks: a state that carries history, dynamics that can be rolled forward, and a forecast whose uncertainty is explicit. Faithful long-horizon rollout and behaviour under regime change remain open, which is exactly what an evaluation contract is for.

Main finding

The evidence has moved from analogy to measurement. Generative market simulators reproduce order-book behaviour and stylised market facts, financial foundation models forecast across instruments from one pretrained model, and deep order-flow models extract signal at several horizons. What a financial world model adds is the layer those pieces lack: a state that carries history, dynamics that can be rolled forward, and an explicit distribution over outcomes rather than a single number. Two difficulties remain genuinely open, faithful long-horizon rollout and behaviour when the regime changes, and both are measurement problems rather than verdicts on the approach.

What separates a real financial world model

  • A financial world model: keeps a latent state of the market, learns how that state moves, and is scored on rollouts against what actually happened.
  • A generative market simulator: reproduces plausible market paths for stress testing and strategy design. Aimed at one venue and one book, the same machinery becomes a forecaster with a counterfactual.
  • A sequence forecaster: maps history to a distribution over the next value, strong at short horizons, with no simulated path behind the number.
  • A chat interface over market data: retrieves and describes, and starts again from nothing with each question. Useful, but it holds no dynamics.

Benchmarks required before any claim

  • Horizon-specific accuracy against a stated baseline: pinball loss for quantiles, directional hit rate, and calibration of the predictive distribution.
  • Economic value after costs: the forecast has to change a decision, and the change has to survive fees, spread, impact and capacity limits.
  • Regime holdout: evaluate on a period whose volatility and liquidity regime was absent from training, then report the degradation rather than hide it.
  • Rollout fidelity: a simulated path should reproduce observed volatility clustering, depth and impact, not only the price level.
  • Stability: accuracy has to survive retraining, data revision and live drift, measured monthly rather than asserted once.

What supports it

  • Generative market simulators conditioned on order-level behaviour reproduce realistic market dynamics and let a strategy be tested against stress scenarios before capital is committed (Li et al., MarS, ICLR 2025).
  • Large-scale generative agent simulators reproduce stylised market facts, including volatility clustering and fat-tailed returns, which gives a substrate for regime experiments rather than only descriptive output (MarketSim, ICML 2026).
  • A foundation model pretrained on price sequences transfers across instruments and horizons without per-market feature engineering (Kronos, 2025), and the 2026 financial world modelling review frames state plus learned dynamics as the organising direction for forecasters of this class (arXiv:2610.09048, October 2026).
  • Deep networks learn universal features of price formation, and order-flow imbalance carries predictive signal at horizons well beyond the shortest (Sirignano & Cont 2019; Kolm et al. 2023).
  • In crypto markets, where the order book and the chain are both observable, order flow carries information about subsequent returns (Journal of Financial Markets, 2026).
  • The OFR's 2026 review finds deep learning improves forecast accuracy most where relationships are non-linear and data are rich, and least where signal-to-noise is thin: the scope has to be argued horizon by horizon (OFR, August 2026).

What this does not claim

  • Rollout error compounds with horizon: a simulated path is informative over the window it was validated on, and has to be refreshed with observed state beyond it.
  • Reflexivity is structural: once a forecast changes behaviour it changes the distribution it was learned on, so adaptation rather than stationarity is the working assumption.
  • Much of the published work reports in-sample or short-horizon results; a market-scale claim still requires an independent out-of-sample record, which no vendor has published.
  • Accuracy is not profitability, and backtests overfit unless the evaluation period is held out before modelling begins.

Planning assumptions used

  • Work starts from one measurable subsystem with a recorded baseline, and widens to a cross-asset market state only after that subsystem passes its own contract.
  • Horizon, metric and failure criteria are written before the first result.
  • Every scenario is stored with its inputs so a decision can be reviewed afterwards.

Open before signature

  • Which recurring decision does the forecast change, and what is its baseline today?
  • What hold-out period and metric will be agreed before evaluation begins?
  • Would a decision actually change if the simulation disagreed with the desk?

Sources

Figures marked illustrative are planning assumptions. This note separates published evidence from project assumptions and does not present a supplier quotation, a valuation or a forecast of returns.