BetaProjection accuracy
Every projection run is kept with the stage it was made at. Once a gameweek finishes, each run is scored against the points that followed, so this page shows how much the forecast sharpens from the early look to the deadline, how it degrades for gameweeks further out, and where the range is honest. Nothing is scored on information the run did not have. Back to the projections → How the numbers are made →
Gameweeks scored
1
GW6–6
MAE at the last stage
2.24
171 player-gameweeks · backtest 2.35
MAE at the early look
2.32
-0.08 by the deadline
Range coverage
87%
P10–P90 should hold about 80%
Likely starters are those the run gave 60%+ to play an hour; positions are among them.
By stage of the week
The next gameweek only, pooled over every scored gameweek. Each stage is the last run made at that stage.
| Stage | n | MAE | Bias | Rank | In range | Model in range | CRPS | Haul Brier | Mode hit |
|---|---|---|---|---|---|---|---|---|---|
| Early look | 202 | 2.32 | -0.00 | 0.08 | 84% | 89% | 1.567 | 0.038 | 27% |
| Odds in | 182 | 2.33 | +0.00 | 0.09 | 86% | 89% | 1.576 | 0.037 | 28% |
| Post press conference | 171 | 2.24 | +0.13 | 0.12 | 87% | 89% | 1.492 | 0.033 | 29% |
| Deadline | 171 | 2.24 | +0.13 | 0.12 | 87% | 89% | 1.492 | 0.033 | 28% |
| Season backtest (likely starters) | 5,264 | 2.35 | – | 0.22 | 88% | – | 1.605 | 0.052 | – |
Weeks ahead
How the forecast for a gameweek degrades the further out it was made. Lead 0 is the next gameweek; lead 2 is a projection made two gameweeks before.
| Lead | n | MAE | Bias | Rank | In range | Model in range | CRPS | Haul Brier | Mode hit |
|---|---|---|---|---|---|---|---|---|---|
| Next gameweek | 726 | 2.28 | +0.06 | 0.10 | 86% | 89% | 1.534 | 0.035 | 28% |
Market inputs on and off
The same players at the same cutoff, projected with and without the Market FDR and bookmaker inputs. Next gameweek, likely starters.
| Variant | n | MAE | Bias | Rank | In range | Model in range | CRPS | Haul Brier | Mode hit |
|---|---|---|---|---|---|---|---|---|---|
| Default (market on) | 726 | 2.28 | +0.06 | 0.10 | 86% | 89% | 1.534 | 0.035 | 28% |
| no-history | 174 | 2.21 | +0.22 | 0.10 | 88% | 89% | 1.461 | 0.028 | 29% |
| no-market | 589 | 2.31 | +0.01 | 0.10 | 84% | 88% | 1.558 | 0.039 | 27% |
By gameweek
MAE of the mean for the next gameweek, by stage. Blank where no run at that stage was stored.
| Gameweek | Actual avg | Early look | Odds in | Post press conference | Deadline |
|---|---|---|---|---|---|
| GW6 | 3.13 | 2.32 | 2.33 | 2.24 | 2.24 |
Reading the numbers
- MAE is the average absolute gap between the projected mean and the points scored. For likely starters a naive guess of their season average comes in around 2.8 on the backtest; the model's 2.35 is the bar to beat.
- Bias is the signed average: above zero and the model promised more than arrived.
- In range is how often the actual score sat between P10 and P90. Near 80% means the ranges are honest; well above it means they are wider than they need to be.
- CRPS scores the whole distribution, not just its centre; it is the number the model's parameters are tuned on, and lower is better.
- A run is only scored if it finished before the deadline of its first gameweek, and only once the public history for that gameweek is in. The stage is the one the run was labelled with at the time.