Towards Financial World Modeling

A trillion-observation market dataset and systematic study of representation learning in financial markets.

Humzah MerchantAlec GuthrieSimon MahnsRandall BalestrieroBradford Levy

University of Chicago, Booth School of Business

A world model plans from a state representation, often for tasks it was never trained on. In markets, that state has to carry expected returns, liquidity, volatility, market-wide conditions and cross-asset relationships. Yet financial representation learning has mostly been judged on one forecasting task over one slice of time. We release Market-1T, nearly one trillion one-second observations of U.S. equities from 2008 to 2025, and use it to compare 18 encoder-training strategies across 32 months of market regimes, as the state layer for world models such as DINO-WM,[1] V-JEPA 2[2] and LeWM.[3]

Three augmentations, how each ranks among 14 SSL methods on clustering and forecasting, and supervised scaling curves — Matching views of the same stock clusters latents by firm and hardly at all by day. At a fixed compute budget, ViT-Tiny beats larger ViTs.

Market-1T

A best-in-class academic dataset spanning 2008–2025 at 1 Hz frequency, released in partnership with Massive.

Monthly rank IC of one encoder recipe on return, volatility change and spread change, 2008 to 2025 — One recipe, retrained on the six months before each month and evaluated on it, across the full span. The switch picks out our 32 evaluation months.

Rigorous Evaluation Protocol

Every method is evaluated on the same 32 months and trained only on the six months before each one. We report rank IC on 15-minute return, volatility change and spread change.

IC vs Sharpe Ratio Decomposition of Variance Decay Over Time Scaling

Annualized Sharpe of a long-short book versus EFQ for three encoders, exiting after 15 minutes or at the close — A Sharpe ratio depends not only on the forecast but on how return forecasts are translated into portfolios and trades, and on the cost and execution model.

Seed × Month With Random ViT Minus Random ViT

Rank IC for 10 seeds in each of 10 evaluation months, as box plots per month — Ten seeds in each of ten months. The month explains 97.2%, 98.9% and 99.9% of the variance; the seed almost none. A single evaluation period says little about a method.

Rank IC by month for the trained seeds beside the untrained Random ViT — The untrained Random ViT rises and falls with the same months, so most of a month's difficulty is shared by every encoder.

Rank IC minus the Random ViT mean, by month, for ten seeds in each month — Subtracting each month's Random ViT mean removes much of the month-to-month swing. Our standard errors are computed this way.

Rank IC relative to a retrained model versus months between training and evaluation — Each model is compared with the same model retrained just before the evaluation month, to control for the intrinsic difficulty of the tasks changing over time. Return holds up; volatility and spread lose about 20% and 10% of their IC over five years.

Rank IC versus training FLOPs for ViT-Tiny, ViT-Small and ViT-Base — At fixed training FLOPs, smaller encoders win on every task: ViT-Tiny beats a ViT-Base trained with 10× the compute. Every point is an annealed WSD[4] checkpoint.

Self-Supervised Learning

We test different data augmentation strategies and pretraining objectives.

Rank IC averaged over 32 months. A supervised arm's own head IC is in parentheses. Red fails to beat the Random ViT; green marks the best two non-supervised arms.
Encoder	Return	Volatility	Spread
Supervised
Return	0.0272 (0.0269)	0.0695	0.1305 (below floor)
Volatility	0.0197	0.0883 (0.0891)	0.1480 (below floor)
Spread	0.0178	0.0762	0.2517 (0.2519)
Multihead	0.0311 (0.0317)	0.0941 (0.0952)	0.2455 (0.2441)
LeJEPA
Same Stock, Diff. View	0.0182	0.0715	0.1672 (below floor)
Time Warping	0.0204	0.0721	0.1691 (below floor)
Gaussian Noising	0.0175	0.0688 (below floor)	0.1625 (below floor)
Cross Stock	0.0169 (below floor)	0.0693 (below floor)	0.1667 (below floor)
C-S, Same Industry	0.0173	0.0692 (below floor)	0.1660 (below floor)
SSL
DINO	0.0097 (below floor)	0.0310 (below floor)	0.0626 (below floor)
BYOL	0.0104 (below floor)	0.0314 (below floor)	0.0686 (below floor)
CPC	0.0184	0.0697	0.1519 (below floor)
I-JEPA	0.0103 (below floor)	0.0256 (below floor)	0.0646 (below floor)
MAE	0.0181	0.0684 (below floor)	0.1435 (below floor)
TS2Vec	0.0205	0.0719	0.1500 (below floor)
CoST	0.0100 (below floor)	0.0339 (below floor)	0.0302 (below floor)
TF-C	0.0095 (below floor)	0.0570 (below floor)	0.1171 (below floor)
TimeMAE	0.0199	0.0706	0.1408 (below floor)
Untrained floor
Random ViT	0.0170	0.0694	0.1692

Organization

We embed views of six firms from three industries within one evaluation month and ask whether nearby embeddings share a partner, a day, a firm or an industry.

Four retrieval tasks: partner matched view, own-day centroid, own-firm centroid, partner firm centroid

Top-1 rate as a multiple of chance / mean percentile rank of the target (lower is better). Green marks the top three trained methods in a column; red is no better than the untrained Random ViT.
Encoder	Partner Matched View	Own-Day Centroid	Own-Firm Centroid	Partner Firm Centroid
Supervised
Return	2.26× / 44.3%	1.76× / 42.8%	1.25× (below floor) / 45.6% (below floor)	1.17× / 47.3% (below floor)
Vol	2.13× / 44.8%	1.70× / 43.2%	1.91× / 33.9%	1.27× / 45.6%
Spread	1.65× (below floor) / 46.3% (below floor)	1.45× (below floor) / 45.5% (below floor)	2.41× / 26.1%	1.15× (below floor) / 47.0% (below floor)
Multihead	2.25× / 44.4%	1.70× / 43.6%	2.35× / 28.6%	1.38× / 43.8%
LeJEPA
Same Stock, Diff. View	1.65× (below floor) / 45.0%	1.31× (below floor) / 46.0% (below floor)	5.36× / 2.5%	1.23× / 45.5%
Time Warping	1.97× / 39.7%	2.01× / 37.7%	1.54× (below floor) / 40.9% (below floor)	1.07× (below floor) / 47.5% (below floor)
Gaussian Noising	2.33× / 41.1%	2.03× / 39.2%	1.59× (below floor) / 38.6% (below floor)	1.12× (below floor) / 47.3% (below floor)
Cross Stock	2.46× / 40.5%	2.09× / 38.4%	1.55× (below floor) / 39.3% (below floor)	1.13× (below floor) / 47.6% (below floor)
C-S, Same Industry	2.58× / 39.7%	2.22× / 37.6%	1.52× (below floor) / 39.9% (below floor)	1.12× (below floor) / 47.3% (below floor)
SSL
DINO	2.28× / 39.6%	2.04× / 37.9%	1.49× (below floor) / 40.5% (below floor)	1.23× / 46.1%
BYOL	2.23× / 41.0%	1.94× / 39.1%	1.72× (below floor) / 36.4% (below floor)	1.19× / 46.4%
CPC	2.07× / 46.2% (below floor)	1.57× / 45.1% (below floor)	1.55× (below floor) / 40.3% (below floor)	1.15× (below floor) / 48.1% (below floor)
I-JEPA	2.24× / 40.5%	1.89× / 39.5%	1.06× (below floor) / 49.0% (below floor)	1.24× / 45.1%
MAE	2.42× / 40.2%	2.10× / 38.5%	1.23× (below floor) / 45.4% (below floor)	1.22× / 46.7% (below floor)
TS2Vec	1.86× / 43.5%	1.47× (below floor) / 44.4%	3.99× / 11.0%	1.37× / 43.2%
CoST	2.13× / 40.6%	2.00× / 38.1%	1.03× (below floor) / 49.6% (below floor)	1.17× / 47.4% (below floor)
TF-C	2.15× / 45.6% (below floor)	1.55× / 45.0% (below floor)	4.14× / 10.6%	1.39× / 44.7%
TimeMAE	3.35× / 38.6%	2.66× / 36.0%	1.32× (below floor) / 43.9% (below floor)	1.33× / 44.6%
Frozen time-series foundation models
Chronos-2	2.24× / 42.5%	1.88× / 39.9%	2.56× / 24.9%	1.07× (below floor) / 48.4% (below floor)
Kronos	2.90× / 41.1%	2.27× / 38.6%	1.43× (below floor) / 42.2% (below floor)	1.40× / 44.2%
TimesFM 3.0	1.95× / 44.2%	1.55× / 43.4%	2.43× / 28.2%	1.22× / 45.4%
Untrained floor
Random ViT	1.82× / 45.6%	1.48× / 44.4%	1.81× / 36.2%	1.16× / 46.6%

Factor Structure

We estimate statistical factors from the month's realized returns[5] and ask whether each firm's loadings can be decoded from its embedding, and whether the embedding's leading directions span the factor subspace.

Factor loadings estimated from realized correlations, then decoded from embeddings

Decoded loadings: mean out-of-sample correlation. Subspace alignment: share of the factor subspace captured. Higher is better. Green marks the top three trained methods in a column; red is no better than the untrained Random ViT.
Encoder	Decoded Loadings	Subspace Alignment
Supervised
Return	0.633	0.202
Vol	0.581	0.207
Spread	0.459 (below floor)	0.161
Multihead	0.582	0.243
LeJEPA
Same Stock, Diff. View	0.391 (below floor)	0.199
Time Warping	0.518	0.218
Gaussian Noising	0.381 (below floor)	0.162
Cross Stock	0.415 (below floor)	0.167
C-S, Same Industry	0.433 (below floor)	0.173
SSL
DINO	0.544	0.190
BYOL	0.444 (below floor)	0.199
CPC	0.475 (below floor)	0.166
I-JEPA	0.608	0.203
MAE	0.547	0.178
TS2Vec	0.464 (below floor)	0.245
CoST	0.711	0.181
TF-C	0.534	0.221
TimeMAE	0.646	0.207
Frozen time-series foundation models
Chronos-2	0.674	0.221
Kronos	0.620	0.184
TimesFM 3.0	0.716	0.194
Untrained floor
Random ViT	0.495	0.151

Summary

How well an encoder forecasts says little about how it organizes the market. Methods that cluster by firm do not cluster by day, and the forecasting and organization rankings are unrelated.

Task Correlations Forecasting vs. Organization

Rank correlation between every pair of evaluation tasks across 19 encoders — Spearman correlation between the 19 encoders' rankings on every pair of tasks; bold is p < 0.05. Own-day and own-firm clustering pull in opposite directions.

Organization rank versus forecasting rank for 19 encoders — Each encoder's mean rank on forecasting against its mean rank on organization. TimeMAE and the supervised multihead trace the frontier.

Citation

Correspondence: bll@uchicago.edu

\cite{merchant2026towards}

@inproceedings{merchant2026towards,
  title={Towards Financial World Modeling},
  author={Merchant, Humzah and Guthrie, Alec and Mahns, Simon and Balestriero, Randall and Levy, Bradford},
  booktitle={The Fortieth Annual Conference on Neural Information Processing Systems},
  year={2026},
  url={https://arxiv.org/abs/2610.09048}
}

References

Gaoyue Zhou and Hengkai Pan and Yann LeCun and Lerrel Pinto. DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning. Forty-second International Conference on Machine Learning, 2025.
Mido Assran and Adrien Bardes and David Fan and Quentin Garrido and others. V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning. arXiv:2506.09985, 2025.
Lucas Maes and Quentin Le Lidec and Damien Scieur and Yann LeCun and Randall Balestriero. LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels. 2026.
Shengding Hu and Yuge Tu and Xu Han and others. MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies. arXiv:2404.06395, 2024.
Markus Pelger. Large-Dimensional Factor Modeling Based on High-Frequency Observations. SSRN, 2018.