Model leaderboard
forecast skill vs. no-change baseline| # | Model | n | Skill | RMSE | MAE | MAPE | Bias | Dir. |
|---|---|---|---|---|---|---|---|---|
| No graded predictions yet | ||||||||
Skill score = 1 − RMSE ÷ baseline RMSE, where the baseline is the naive no-change forecast (price stays at its reference value). A positive skill score means the model genuinely beats persistence; 0 ties it; negative is worse. Click any column to re-rank.
Forecast quality
interval calibrationTrade profitability
| Model | Hours | Repriced | Shorts | Hit | Avg net | Profit | Sharpe | t | Flip profit |
|---|---|---|---|---|---|---|---|---|---|
| No graded min/max bands yet | |||||||||
Contrarian-short strategy, tracked live and out-of-sample. Backtests over the first weeks of data found the bands' only significant signal is inverted: hours where a model is confidently bullish — band midpoint at least τ half-widths above the market (z ≥ τ) — averaged strongly negative moves. The rule charted above: short at the hour start whenever z ≥ τ, exit at the hour end, net of a round trip (2 × cost). Hours are de-overlapped (models post every 5 minutes; only the freshest band per clock hour counts, so t-stats aren't inflated 12×) and entries are repriced to the nearest market tick (stored reference prices can sit several bps off-market). The week the rule was found it returned +6.2% with t = 2.69 on 17 trades — small sample, found in-sample. This panel is the out-of-sample test: trust it when t stays ≥ 2 as trades accumulate; treat it as dead if profit decays toward fees. Both knobs re-score live.