Towards Financial World Modeling

The gist Building a world model requires a state representation useful for planning and decision-making—potentially over tasks unknown at training time

Market-1T Dataset

The paper introduces Market-1T, which is a dataset containing nearly one trillion observations across U.S. equities from 2008 to 2025 at 1 Hz resolution This dataset combines a longer historical span with substantially finer temporal resolution than is typical of public financial representation-learning benchmarks The dataset is constructed from consolidated U.S. equity quote and trade data distributed by Massive The underlying data consist of National Best Bid and Offer (NBBO) updates and consolidated trade reports Each one-second observation contains the same 11 features: Order Book State, Price Summary, and Activity

Rigorous Evaluation Protocol

The researchers developed a rigorous evaluation protocol to test the sensitivity of their methods This protocol closely examines the sensitivity of the evaluation protocol to probe sample size, market regime, model scale, representation age, and the portfolio-construction choices used to translate forecasts into trades They evaluate every method on the same 32 months sampled across 2008–2025 The evaluation protocol involves training models using only the preceding six months of data for each evaluation month They standardize architecture, compute, preprocessing, and probe fitting across methods wherever possible

Exploration of SSL in Financial Markets

The study systematically compared 14 self-supervised encoder-training strategies under a common architecture and evaluation protocol The goal was to ask whether learning preserves information about future returns, volatility, and liquidity The results showed that encoders with similar predictive performance can organize market state very differently Predictive performance varies substantially across market regimes, making long-horizon evaluation essential

Key Findings on Representation Quality

The paper concludes that the resulting picture is not captured by a single notion of representation quality, as predictive performance varies substantially across market regimes Representations that perform well on forecasting need not be those that best preserve economically meaningful latent organization This suggests a useful financial state representation must preserve more than what is needed for next-return prediction alone

Latent Space Organization

The evaluation of latent space organization revealed that different objectives induce distinct geometries Some methods focus on persistent firm identity while others focus on shared market state For example, the model that is trained to see views of the same stock at different times learns to organize them using the bid and ask size information

Metric Preference

The authors prefer the information coefficient (IC) over the Sharpe ratio as their primary predictive metric They state that IC is much more closely related to performance after implementation choices are averaged away than it is within any particular backtest

Open Access and Reproducibility

The dataset and pre-trained encoders are available at https://huggingface.co/finai-lab The code is available at https://github.com/fin-ai-lab/tfwm

Limitations

The evaluation of latent space organization captures largely distinct properties of a representation from predictive performance Predictive performance varies substantially across market regimes, making evaluation on a single recent period unreliable The paper notes that the evaluations are limited in scope The authors discuss that the results are dependent on implicit assumptions

Conclusion

A useful financial state representation must therefore preserve more than what is needed for next-return prediction alone This matters for world modeling because a model planning an action—for example, executing a trade—may need to reason jointly about expected returns, liquidity, volatility, market-wide conditions, and cross-asset relationships in order to anticipate both market evolution and the consequences of that action<ref:2610.09048#pg6>

The paper is concerned with building a foundation for training and evaluating financial market representations in support of world models such as DINO-WM, V-JEPA 2, and LeWM The study addresses the need for a state representation that enables planning and decision-making under a highly stochastic environment<ref:2610.

1. textbfImprove World Model Reasoning via Multi-faceted State Preservation: The improved system can reason jointly about expected returns, liquidity, volatility, market-wide conditions, and cross-asset relationships because it learns representations that do not rely on a single notion of quality but instead preserve predictive information about future return, volatility change [1], and bid–ask spread change.

2. textbfEnhance Model Robustness Across Regimes: The system will be better at adapting to changing market conditions because the evaluation protocol is designed to address the fact that predictive performance varies substantially across market regimes, moving beyond evaluation on narrow slices of time by using a long-horizon protocol sampling 32 months throughout 2008–2025.

3. textbf Optimize Representation Learning for Specific Structures: The improved system can learn representations tailored to specific economic structures; for instance, it can better capture shared market state when trained with the Cross-Stock augmentation, or focus on persistent firm identity when using the Same Stock Views view.

Building a world model requires a state representation useful for planning and decision-making---potentially over tasks unknown at training time. In the context of financial markets, planning and decision-making may require a model to reason about market-wide conditions, asset-specific expected returns, liquidity, volatility, and cross-asset relationships. Yet financial representation learning has largely been evaluated on individual predictive tasks, oftentimes on a single time period using comparatively narrow datasets. We address this through three primary contributions. First, we introduce Market-1T, a dataset containing nearly one trillion observations across U.S. equities from 2008 to 2025 at 1 Hz resolution. Second, we develop and implement a rigorous evaluation protocol. Third, we conduct a systematic large-scale study of financial representation learning, comparing 18 encoder-training strategies across nearly two decades of market regimes. We evaluate learned representations both by their predictive utility on common finance tasks and through probes of latent structure. We find that encoders with similar predictive performance can organize market state very differently. Collectively, we establish a foundation for training and evaluating financial market representations in support of world models such as DINO-WM, V-JEPA 2, and LeWM.