Machine-readable dump of the full research trail (ours vs TWC vs market, every city-day). Built for copy-paste off-box backup — select a box, copy, paste into a Google Doc / Sheet whenever you like. Nothing here is pretty on purpose. Auto-refreshes from the same hourly collector as the dashboard.
One row per city per day. status=settled has TWC truth + P&L; status=pending is frozen-9am ours/market awaiting settlement. Columns: date, city, our_center, our_pick, mkt_ev, mkt_fav, disagree(1/0), twc_high, win_bucket, our_off, mkt_off, our_hit(1/0), mkt_hit(1/0), our_price(¢), pnl_c(¢), status.
Everything: the flat rows above, PLUS the untouched per-day source (raw_boards = our 9am freeze + market lines; raw_settlements = TWC hourly + derived high). This is the complete backup — if the box dies, this JSON reconstructs the whole trail.
Fixed out-of-sample benchmark of the corrector, not a live daily number. Re-runs the same 80/20 split on a frozen 2yr training set (through 2026-09-13), so values are stable until we retrain. Training/validation ground truth here = IEM ASOS station daily high — the only source with a 2yr archive to learn from. (Live daily performance vs The Weather Company (TWC), the number Kalshi pays on, is on the main dashboard — different truth, different job.) raw MAE = public multi-model avg error vs station · corr MAE = our bias-corrected error · ERA5 err = how far the old reanalysis target was from the station (why we switched). Moved here from the main dashboard — it's reference material, not something that changes day to day.
~/.openclaw/workspace/projects/weather-quant/; this page is a copyable mirror.