Performance
Batter Hits is running below the rate it converts without us (-7.7 pts, n=53).
Overall hit rate is 59.6%, +13.6 pts over baseline — down 1.5 points on the previous 7 days, which ran +12.3 pts.
Performance summary
The model is showing meaningful predictive signal over this period — +13.6 pts over baseline on 146 graded predictions.
What's working
Every market ranked by the edge it holds over the rate its own event happens without us — not by how often it converts. The tables further down read served predictions only, so the “all” figure beside each edge is the check that a market which never clears the floor is not mistaken for one that does not work.
Trust board
The page’s verdict, from edge over baseline, sample size and recent trend together — never from hit rate alone. A tier with nothing in it is not shown.
- Batter Hits-7.7 pts n=53
- SP Strikeouts+43.4 pts n=14improving, +39.7 pts
- Points+35.3 pts n=24
- SP Hits Allowed+21.8 pts n=11
- Rebounds+19.8 pts n=20
- Batter Ks+18.2 pts n=5
- Assists+11.4 pts n=19cooling, -18.1 pts
Score calibration
The scoring system is worth having only if a higher score means a better outcome. Each band is measured against the base rate of the props it actually holds — the top band is almost purely 1+ hit, so a raw-rate reading of this table would credit it for carrying a common event.
Score is directionally useful, but not perfectly monotonic — 99–100 runs +30.7 pts above 70–74 over base, with 1 of 2 steps going the wrong way.
Predictions scored 90+ outperform everything below them by 31.3 pts over base (n=56).
The 80–84, 85–89, 90–94 and 95–98 bands remain small; read them as an observation, not a rate.
Where the model has edge
Segments by market, ranked by lift over their own base rate. Anything under the minimum sample is separated and labelled rather than ranked alongside — a high percentage on twenty predictions is not a discovered edge.
Over vs Under
Under signals are outperforming Over by 20.2 points of lift (+31.8 pts against +11.6 pts). The under side has only 14 graded predictions, so treat the gap as an observation.
Consistency
Performance has been stable around 60%–63% across every window with a real sample.
Market pulse
Each market’s recent slates against its own period average. The trend column is the row’s answer; the grid behind it fades with age.
By month
Model version performance
3 of 7 current model versions are outperforming the versions they replaced; MLB Batter Hits, MLB SP Strikeouts, WNBA Points and WNBA Rebounds are behind. WNBA Rebounds and MLB Batter Ks rest on a sample too small to call.
Methodology
- Edge over baseline, never a raw hit rate
- Every “vs base” figure is the observed hit rate minus how often that exact prop’s event happens on its own, for the population that could plausibly have been offered it. A batter getting 1+ hit happens ~61% of the time unprompted; a starter striking out 8+ happens ~13% of the time. Against one shared average the rare-event markets look bad no matter how well they are picked.
- Opportunity Score is a ranking signal, not a probability
- 70 is the publication floor. A 93 is not “93% likely” — it is an estimate of edge over the market’s own base rate. Observed hit rates on this page are outcomes, not forecasts.
- Sample size is a headline, not a footnote
- Under 30 graded predictions a segment is separated and marked rather than ranked. Predictions inside one slate are correlated, so a run of good days is less independent evidence than its count suggests.
- Cohorts
- All qualifying is every published prediction scoring 70+. Featured is the eight carried on Today. Other qualifying is the remainder — the split exists to check that ranking, not just scoring, is doing something.
- Calibration reads the current engine only
- A superseded scorer’s calibration is not a fact about the one running today. Pooling engines once made the top band look anti-predictive when it was not. Per-version records stay in the model table above.
- What is not here
- No odds are consumed anywhere on this page. Total bases and batter walks are excluded — both were retired for being unservable rather than for being wrong. Voids and pending predictions are excluded from every hit rate.