Live Model
2026-W36-original-defaults
Restored original hand-tuned FPL contribution coefficients from the pre-tuning model.
6 Sept 2026, 01:00Last live update9Weighted factors
FPL Model Transparency
FPL Model: Latest Weights, Updates and Validation
See what TallyKind's Fantasy Premier League model uses, how weekly evidence is reviewed, and whether the latest tuning run improved the validation metrics enough to change the live model.
Latest Validation
2026-W37
Latest validation was not applied: No clean pre-gameweek snapshot was available.
8 Sept 2026, 02:09Report generated5,500Player/GW rows
Model Inputs
Advanced signals, kept explainable
The player score combines a prior points baseline with expected minutes, attacking involvement, clean-sheet context, defensive/save contribution, bonus-points signal, set pieces, market signal and conceded-risk adjustment.
Each player table exposes a compact "why this score" breakdown, so the model is not just a single unexplained projection.
Learning Loop
Bayesian-style cautious updating
The live weights act as the prior. Weekly evidence comes from public FPL outcomes. Suggested changes are treated as evidence, not automatically trusted, and only pass if the validation gates improve enough.
This is Bayesian-style model discipline rather than a claim that every player projection is a full posterior probability distribution.
Validation
Latest error parity plot
This compares live model RMSE against the suggested tuned weights from the latest validation report. Points below the diagonal improved in that run; points above the diagonal worsened.
Error parity: live vs suggested
Lower suggested RMSE is better.
Overall Result
Suggested weights improved this test
+0.069RMSE improvement+0.147MAE improvement+0.019Correlation change
The latest report was based on current public aggregates from the Fantasy Premier League public API. It was not applied to the live model unless all quality gates passed.
Accuracy Stats
Current versus suggested weights
| Group | Rows | Live RMSE | Suggested RMSE | Live MAE | Suggested MAE | Correlation |
|---|---|---|---|---|---|---|
| Overall | 5,500 | 3.135 | 3.066 | 2.703 | 2.556 | 0.454 |
| DEF | 1,565 | 2.528 | 2.515 | 2.163 | 2.083 | 0.293 |
| FWD | 1,278 | 4.052 | 4.059 | 3.623 | 3.565 | 0.759 |
| GKP | 500 | 1.891 | 1.980 | 1.273 | 1.097 | 0.254 |
| MID | 2,157 | 3.141 | 2.953 | 2.881 | 2.640 | 0.212 |
Latest Weights
Live model and latest suggested tuning
| Factor | Live weight | Suggested weight | Suggested delta |
|---|---|---|---|
| Baseline | 0.450 | 0.370 | -0.080 |
| Appearance | 0.550 | 0.390 | -0.160 |
| Attack | 0.550 | 0.790 | +0.240 |
| Clean Sheet | 0.550 | 0.550 | 0 |
| Defensive Save | 0.550 | 0.550 | 0 |
| Bonus Bps | 0.550 | 0.950 | +0.400 |
| Set Pieces | 0.550 | 0 | -0.550 |
| Market | 0.550 | 0.750 | +0.200 |
| Conceded Risk | 0.550 | 0.550 | 0 |
Model Notes
What this page means
Why was the latest tuning not applied?
No clean pre-gameweek snapshot was available.
What will improve over time?
The weekly process captures a pre-GW snapshot, then compares it with actual points once the GW is complete. More clean snapshots should make the validation history more useful than a one-off report.
What data is stored?
The report stores public player IDs, feature vectors and aggregate metrics. Sampled manager IDs are used for evaluation but are not written to the report or snapshot.
Use the model
Run the projections with context
Use the live FPL tools, then open player score breakdowns when you want to see why a player scored well.