Trillion-observation finance benchmark: linear ridge probe beats 11 of 14 self-supervised encoders
burny_tech · x · 2026-10-24
A Brown/Chicago/JHU team presents "Towards Financial World Modeling" (NeurIPS 2026), introducing Market-1T: 1 trillion observations of US equities from 2005–2025 at 1Hz resolution, plus a rigorous evaluation protocol.
Key findings:
- A linear ridge probe on raw input achieves 0.0196 return IC, beating 11 of 14 self-supervised encoders; an untrained Random ViT scores 0.0170
- They compare 18 encoder-training strategies across two decades of market regimes and show return prediction and latent market organization are decoupled (rho = -0.19)
- View selection and time resolution strongly shape latent clustering and downstream volatility/spread-change prediction
The work aims to establish a foundation for evaluating financial market representations for world models like DINO-WM, V-JEPA 2 and LeWM. Data, models and code are open-sourced.
More from Venture
- Stripe reportedly buying OpenRouter for ~$7.5B as AI usage billing becomes the prize — FinanceYF5 · 2026-10-24
- Anthropic reported an $80B net loss in 1H, but $72B came from convertible-note accounting charges — Beth_Kindig · 2026-10-24
- GitHub pushes up 4.9x in a year — the 'GitHub killer' may be GitHub itself — FinanceYF5 · 2026-10-24
- Salesforce to Require Registration of External AI Agents, Charge per Data Access — verrsane · 2026-10-24
- Going-global dev: set up account recovery for Google, Apple, Stripe before it bites you — sven_ai · 2026-10-24
- AI transforms content supply economics but not what consumers actually want — illscience · 2026-10-24