FPL projection accuracy & track record.
3 gameweeks settled
Model updated
Across 3 settled gameweeks and 960 player-projections, Metis projections miss by 1.89 points per player on average, with an overall bias of +0.10 (actual minus projected) and a rank correlation of 0.57 between projected and actual points.
The projection of record for each gameweek is the last nightly publish before that deadline — never a later one. Actual points are joined from the official feed once the gameweek finishes, and every miss is published exactly like every hit.
Mean abs error
1.89
points per player
Bias
+0.10
actual − projected
Rank correlation
0.57
Spearman, projected vs actual
Top-20 projected
6.0 → 4.3
projected → actual, mean
By gameweek
Every settled gameweek
| GW | Players | MAE | Bias | RMSE | Spearman | Top 20 proj → act | Fantasy 11 proj → act |
|---|---|---|---|---|---|---|---|
| GW3locked Fri 4 Sept | 320 | 1.80 | +0.04 | 2.57 | 0.55 | 5.2 → 3.5 | 5.1 → 3.8 |
| GW4locked Sat 12 Sept | 320 | 2.05 | +0.04 | 2.94 | 0.54 | 5.4 → 5.3 | 4.9 → 4.3 |
| GW5locked Fri 18 Sept | 320 | 1.82 | +0.21 | 2.75 | 0.62 | 5.4 → 4.0 | 5.3 → 4.1 |
Calibration
Projected points vs what happened
Players grouped by projected points (all settled gameweeks pooled). A well-calibrated model shows the actual mean tracking the projected mean in every bin.
| Segment | Players | Projected | Actual | Bias | MAE |
|---|---|---|---|---|---|
| 0–1 xPts | 220 | 0.32 | 0.53 | +0.21 | 0.47 |
| 1–2 xPts | 114 | 1.52 | 1.60 | +0.09 | 1.40 |
| 2–3 xPts | 242 | 2.55 | 2.57 | +0.02 | 2.05 |
| 3–4 xPts | 246 | 3.46 | 3.61 | +0.15 | 2.45 |
| 4–5 xPts | 103 | 4.37 | 4.89 | +0.53 | 3.26 |
| 5–6 xPts | 29 | 5.46 | 3.86 | -1.60 | 3.20 |
| 6–8 xPts | 5 | 6.54 | 5.00 | -1.54 | 4.36 |
| 8+ xPts | 1 | 8.32 | 6.00 | -2.32 | 2.32 |
Start probability
Did they actually play?
| Start probability bin | Players | Mean start% | Appeared | Played 60+ |
|---|---|---|---|---|
| 0–25% | 252 | 5% | 33% | 9% |
| 25–50% | 85 | 39% | 78% | 33% |
| 50–70% | 93 | 60% | 91% | 77% |
| 70–85% | 104 | 77% | 90% | 74% |
| 85–95% | 159 | 92% | 92% | 81% |
| 95–100% | 267 | 97% | 97% | 92% |
"Played 60+" is the closest official proxy for a start; the feed does not publish lineups retrospectively.
Breakdown
By position and price
| Segment | Players | Projected | Actual | Bias | MAE |
|---|---|---|---|---|---|
| GK | 120 | 1.64 | 2.00 | +0.36 | 1.35 |
| DEF | 332 | 2.47 | 2.76 | +0.29 | 2.18 |
| MID | 411 | 2.59 | 2.52 | -0.08 | 1.73 |
| FWD | 97 | 2.88 | 2.74 | -0.14 | 2.23 |
| Segment | Players | Projected | Actual | Bias | MAE |
|---|---|---|---|---|---|
| ≤£5.0m | 482 | 2.07 | 2.23 | +0.16 | 1.69 |
| £5.1–7.0m | 417 | 2.68 | 2.71 | +0.02 | 2.00 |
| £7.1–9.5m | 52 | 3.71 | 4.12 | +0.41 | 2.69 |
| £9.6m+ | 9 | 5.83 | 4.22 | -1.61 | 2.78 |
Squads
Fantasy 11 and Metis FC, projected vs actual
| GW | Fantasy 11 proj | actual | Metis FC proj | actual | GW average |
|---|---|---|---|---|---|
| GW5 | 66.9 | 51 | 64.1 | 48 | 48 |
| GW4 | 61.2 | 59 | 60.9 | 53 | 69 |
| GW3 | 63.4 | 56 | 62.5 | 45 | 51 |
| GW2 | — | — | 53.3 | 101 | 81 |
| GW1 | — | — | — | 43 | 50 |
Method
How this is scored
- Every nightly publish snapshots all projected players. The projection of record for a gameweek is the last snapshot before its deadline; later snapshots are never scored.
- Actual points and minutes come from the official FPL event feed once the gameweek is flagged finished. Nothing is re-projected after the fact.
- MAE is the mean absolute error in points per player; bias is actual minus projected (positive means the model under-projected); Spearman is the rank correlation between projected and actual points over the whole pool.
- Expected points are a mean over a right-skewed distribution: a single player's actual score is usually far from their xPts, so per-player error is large by construction. Calibration — whether groups of players score what they were projected on average — is the fairer test.
- Players outside the projected pool (not published) are not scored either way.
Last updated 23 Sep · 04:07 UTC