The blend is logit(p) = λ·logit(model) + (1−λ)·logit(market).
At 1.00 the market is ignored. At 0.22 — the weight the regression measured —
the model only nudges the price.
How far the blend has to beat the market price before the play counts. Raising it is the usual instinct when a strategy is losing — try it and watch what happens to the win rate.
- Plays clearing the bar
- —
- Blend predicted
- —average
- Actually happened
- —win rate
- Flat-stake ROI
- —
- Largest edge found
- —across all plays
What you're looking at
Every one of these 10,928 predictions was logged before its game started, with the sportsbook's price recorded alongside. Nothing here is refit to the outcome — the slider only changes how the two existing forecasts are combined.
At λ = 1.00 the model ignores the market entirely and finds edges everywhere. The observed line sags below the predicted line and keeps sagging as disagreement grows: the more confident the model, the more wrong it is.
At λ = 0.22 the two lines converge across most of the range and almost nothing clears a 5% bar. That's not the model being switched off — a regression of outcomes on both forecasts says the model carries real information the closing price doesn't (β = 0.231, z = 2.20). It's the model being weighted at what it earned instead of what I assumed.
Raising the edge threshold makes it worse, not better, which is the counterintuitive part. The plays with the biggest disagreement aren't the best opportunities — they're the ones where the model is most wrong.