Driving · 2026-10-08
Towards Financial World Modeling
Humzah Merchant, Alec Guthrie, Simon Mahns, Randall Balestriero, Bradford Levy
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors study state representations for “world models” in financial markets. They introduce Market-1T, a publicly available dataset with nearly one trillion one-second aggregates across U.S. equities (2008–2025), and propose a 32-month evaluation protocol. Using 18 encoder-training strategies, they evaluate predictive utility (rank IC) and latent organization (retrieval/probing of structure), showing that similar predictive performance can correspond to different latent organization.
Why it matters
If representations are to support planning and decision-making, they must preserve the right information for forecasting and for capturing economically meaningful structure. The paper provides a large-scale dataset and a multi-regime evaluation protocol to test financial representation learning under changing market conditions, and shows that representation quality is multidimensional rather than determined by prediction alone.
Method
- Introduce Market-1T: nearly one trillion 1 Hz quote-and-trade aggregates across U.S.-traded equities (2008–2025), with benchmarks constructed per month to avoid survivorship/look-ahead bias.
- Develop and apply a rigorous evaluation protocol that samples 32 months across 2008–2025; train on preceding data and select hyperparameters on separate tuning months; use standardized architectures/preprocessing and probe fitting across methods.
- Systematically compare 18 encoder-training strategies (supervised, self-supervised, and pretrained) and evaluate them via predictive rank IC at a 15-minute horizon and latent structure tests (effective rank, retrieval/proximity by day/firm/industry, and latent statistical factors).
Limitation
The paper reports that predictive performance varies substantially across market regimes, so evaluation on narrow time slices can be biased; this motivates their long-horizon protocol. (The provided text does not state other explicit limitations.)
Abstract (from arXiv)
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.
Related papers
- STOCK-JEPA: Prior-Anchored Latent Revision Representation Learning in Equity Markets
- Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books
- Market-Conditioned JEPA: Learning from Unpredictable Market Shocks
- Pythia: Toward Foundation World Models for Multimodal Time Series
- On the Divergence of Accuracy and Mechanism Consistency in Time Series World Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-06 · Indexed by Pub-AI: 2026-10-08 UTC
Towards Financial World Modeling
Preprint posted: 2026-10-06 · Original record ↗
- arxiv
- 2610.09048
Explicit source versions (1)
- v2 · 2026-10-13
- Explicit source version detected · v2
Source date: 2026-10-13 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:id explicit version (official metadata)
- Preprint available
Source date: 2026-10-06 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.
