Research map / Label-efficient and robust learning / Self-supervised and contrastive

Towards Financial World Modeling

NeurIPS 2026 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Invariance-based self-supervised learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Market-1T, a ~1-trillion-observation 1 Hz dataset of U.S. equity quotes and trades (2008–2025), and systematically compares 18 supervised, SSL, and pretrained encoders under a multi-regime evaluation protocol for financial world-model representations.

Problem

Financial representation learning has been evaluated on narrow predictive tasks over single time periods and narrow datasets, leaving unclear what state representations are useful for financial world models across market regimes.

Approach

Builds a large, long-span, 1 Hz U.S. equity dataset and a rigorous multi-month evaluation protocol, then compares supervised, self-supervised (LeJEPA-style with financial augmentations such as time warping and cross-stock views), and frozen TSFM encoders on both predictive IC and latent-structure probes (organization, effective rank, statistical factors).

Key results

• Supervised multihead model achieves rank IC 0.0311 on return and 0.2455 on spread, outperforming SSL methods, which mostly fail to beat a random ViT on spread prediction.
• ViT-Tiny outperforms a ViT-Base trained with 10x the FLOPs on return, volatility, and spread tasks at fixed compute.
• Predictive rank and latent organization rankings are weakly related (Spearman rho = -0.19); supervised encoders have very low RankMe (about 2.8-3.7) versus 11-17 for many LeJEPA/SSL variants.

Surprising findings

• A random ViT with a well-fit probe is a nontrivial baseline, and most SSL methods give only modest gains or underperform it on spread change.
• Encoders with similar predictive IC can organize market state very differently, so predictive performance alone is not a sufficient evaluation.

Limitations

• Full public release of Market-1T is pending with Massive; the released version is a code path to assemble it manually.
• Evaluated only on a liquid U.S. common-stock universe and at a small model scale (ViT-Tiny to ViT-Base).
• Forecasting and latent-structure results depend on the chosen evaluation window and probe sample size, with large variance across months.

Experimental setup

Datasets: Market-1T

Benchmarks: Market-1T 15-minute rank IC forecasting (return, volatility, spread) latent organization tests (T1-T4) statistical factor probes (F1, F2) RankMe

Metrics: rank information coefficient top-1 rate multiple of chance percentile rank factor decoding score RankMe Sharpe ratio

Baselines: DINO BYOL CPC I-JEPA MAE TS2Vec CoST TF-C TimeMAE Chronos-2 Kronos TimesFM 3.0

Builds on: LeJEPA DINO BYOL I-JEPA MAE

Closest papers by meaning

• Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context (AAAI 2024 · similarity 0.87)
• A Set-Sequence Model for Time Series (NeurIPS 2026 · similarity 0.85)
• R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization (NeurIPS 2025 · similarity 0.85)
• Retrieval-Augmented Diffusion Modeling for Stochastic Discount Factor Portfolios (NeurIPS 2026 · similarity 0.85)
• MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model (ICLR 2025 · similarity 0.84)
• Generating Financial Time Series by Matching Random Convolutional Features (NeurIPS 2026 · similarity 0.84)