Long-only sleeve trading officer / director / 10%+ owner open-market purchases on small & mid caps. The academic edge: insider buys carry ~10-15% 90-day excess return. No take-profit — held a full 60 days with a wide protective stop.
A mechanical sleeve that ranks the universe by 12-1 month momentum, holds the top 20 equal-weighted, and rebalances on the 15th of each month. No stops, no targets — exits are purely time-based at the next rebalance. Excluded from headline P&L so the news + PEAD + Form 4 numbers stay clean.
Each dot is one 5pp predicted-probability bucket — hover for the trade count. The diagonal is perfect calibration: when the model says 60%, trades in that bucket should win ~60%. The shaded band is ±10pp (the gate to flip CALIBRATOR_GATE_MODE off "off"). Dot opacity tracks sample count — faded dots are sparse buckets you should ignore. Scores stored at signal-time, so each pair is naturally out-of-sample — no training leak.
Each dot is one closed trade. X = the return Kairos predicted at signal-time (its EV per share as a % of entry price). Y = the actual % return. Both axes are % so a $500 stock and a $20 stock are comparable — raw $/share let the single priciest name dominate the line. The diagonal is perfect calibration; the dashed line is the trades' best-fit slope (robust Theil-Sen, so one outlier can't flip its sign). Close to 1.0 means predicted and realized moves match; below 1 means the model overstates magnitudes. Check R² too: near zero means the line carries no signal yet, whatever its slope.
Rather than guess which recalibration is best, all three are scored out-of-sample every night and logged here. Lower line = better calibrated (predicted win-rate closer to reality). Once one method beats raw's median by ≥2pp over ~45 nights, it's selected automatically; until then selection stays on raw (current behavior) and nothing changes live. Isotonic usually overtakes Platt once there are a few hundred trades.