Jane Street's 12GB contest data hides a trap: right forecasts can score below zero
Roger_M_Taylor · x · 2026-10-18
Jane Street released 12GB of real production trading data with a $120K prize, and its scoring rule contains a trap: predicting nothing scores zero, but a correct yet overconfident model scores below zero. A paper derives the closed form R² = ρ²(2m − m²): betting at m=1 maximizes score, m=2 wipes it out, and beyond that you lose even when right — the same curve as Kelly criterion sizing. LTCM's two Nobel laureates ran 25x leverage and lost $4.6B in four months by being right at the wrong size. The paper includes a validation protocol and a copy-ready research prompt for turning AI into a backtesting research desk.
More from Venture
- Checking 56 'AI killed X' claims: 6 gone, 22 shrunk, 13 grew — Mastbubbles · 2026-10-18
- Dev ships AI-written game to 3 portals in an afternoon, Opus 5.5 did all the coding — Roger_M_Taylor · 2026-10-18
- Marc Lou and Cognition host Dubai indie hackers meetup, $200 Devin credits for attendees — marclou · 2026-10-18
- BofA: AI GPUs run near 100% utilization, rental rates up 12%-37% YTD — not dotcom glut — McDonaghMatthew · 2026-10-18
- Trader claims $418K PnL in 147 days from a Grok-built Polymarket quant bot — Roger_M_Taylor · 2026-10-18
- Pomp on competing with trillion-dollar giants and hiring AI-native talent — petergyang · 2026-10-18