A point-in-time forecasting system that locks a probability before the outcome is knowable, then scores itself against reality.
forecast.lumixraku.org · live demo, no signup
Python · SQLite (4 tables) · Cloudflare Workers + D1 · gated LLM agent harness | source
No look-ahead, by construction. Every factor is standardized on a strictly recursive expanding window using data up to session t only. Every tool output, headline and citation passes an as_of gate that raises on post-cutoff data instead of silently clipping rows.
Measured, not asserted. 629 signals since 2002 — 66% precision against 59% for a random day (+7.0 pp lift), 67% out-of-sample since 2015. Re-settled on the tradable QQQ ETF, the same signal holds at +7.0 pp, with 1.4% flat-call disagreement across 18 years.
One signal, three markets. The US close lands at 04:00 Beijing time — hours before the Shanghai open — so the unmodified US signal is mapped onto the next 000001.SS session: 59.9% precision vs a 53.1% baseline, +6.8 pp over 476 signals.
It grades its own agents. A framework-free, 8-step tool-using forecasting agent, gated by the same as_of check, benchmarked across DeepSeek and Qwen on the shared intersection of answered questions and scored with Brier / LogLoss / ECE.
Ships as a dashboard. Local no-build server on 4 flat SQLite tables (107k bars · 127k time-gated headlines · 13.7k signals), with a live factor switchboard that recomputes precision and out-of-sample hit rates in the browser — then deploys to Cloudflare Workers + D1.





