PitchProphet Accuracy Tracker
Every projection is logged and auto-graded against the final box score. No cherry-picking — this is the full record across 455 graded starts (1818 decided calls · 2 pushes).
By Metric (all-time)
Skill lives in MAE (average miss) and bias (systematic over/under) — lower and nearer zero is better. The W-L / Hit % columns are directional-vs-our-own-line with exact ties scored as pushes (P, excluded from the record) — shown for context, not as a skill score.
| Metric | MAE | Bias | Record (W-L·P) | Hit % |
|---|---|---|---|---|
| Strikeouts (over) | 2.12 | +0.53 | 193-262 | 42% |
| Outs / IP (over) | 3.91 | +1.92 | 182-272 · 1P | 40% |
| Earned Runs (under) | 1.59 | +0.01 | 256-198 · 1P | 56% |
| Walks (under) | 1.08 | +0.31 | 281-174 | 62% |
Self-Calibration
The engine re-fits its own weights from graded results under bounded guardrails, then syncs the update to the shared ledger. This is that loop's live state — so a weight that drifts far from the calibrated baseline shows up here instead of silently skewing projections.
| Weight | Baseline | Active | Δ vs base | Status |
|---|---|---|---|---|
| split_k_sensitivity | 0.5 | 0.749 | +50% | watch |
| park_k_sensitivity | 0.15 | 0.082 | -45% | watch |
| split_er_sensitivity | 0.4 | 0.220 | -45% | watch |
| split_bb_sensitivity | 0.3 | 0.167 | -44% | watch |
| park_er_sensitivity | 0.7 | 0.404 | -42% | watch |
| sfm_penalty_per_rv | 0.03 | 0.022 | -26% | watch |
0 railed / 6 watch — flagged for a human look. Every weight is still clamped in the projection math, so a flag is a review trigger, not a live error.
Recent Form
How to read this: the directional hit rate scores each start against our own projection (K/Outs win when the pitcher tops it; ER/BB win when under it), and an exact tie is a push — excluded from the record, sportsbook-style, so projections that land on round numbers can't pad the rate. Even so it is not a pure skill score: its baseline still shifts with each metric's distribution shape. Judge the model by MAE (average miss) and bias (systematic over/under) above. Overall MAE across all metric calls: 2.18.