Oxford's LOB-Impact benchmark: generative order book models emerge square-root market impact
oxford-lob · hf · 2026-10-30
Oxford's LOB team released the LOB-Impact benchmark on Hugging Face, testing whether neural generative limit-order-book models reproduce real market impact. Using one month of LOBSTER data, it runs three phases (free generation of 12k–73k messages, metaorder build-up with 100 child orders at 10% participation in visible/blind runs, and relaxation) against five claims including the square-root law (δ≈0.5, Y≈0.5–1). Key finding: models trained purely on next-token prediction exhibit emergent market impact under metaorder injection, though relaxation decay remains unobserved due to limited context windows.
More from Research
- Paper Argues One-Click AI Video Will Break Recommendation Pools, Proposes BoLe System — BreadStudio · 2026-10-30
- PersonaCUABench shows computer-use agents fail real user personas across 74 tasks — Haoming Li · 2026-10-30
- PreView: Training-Free Preprocessor Boosts GUI Grounding via Zoomed Views and Cues — Wenjun Hou · 2026-10-30
- SWE-Verifier: a 4B model lifts coding agents ~1.7pp on SWE-Bench at 9–13% extra cost — radamihalcea · 2026-10-30
- Real citations can still be wrong evidence: the hidden failure mode of AI literature answers — yi111 · 2026-10-30
- BindCraft protein design on an M5 Max MacBook rivals RTX 5090, benchmarks show — AllThingsApx · 2026-10-30