Model versions

v0.6 is now live. Previous versions stay visible here so you can see how the model is changing — and judge for yourself whether each new release earns the version bump. Only challengers that beat the current version get promoted.

v0.4 (backtest)

Simulated

XGBoost stacker on Dixon-Coles base, with cross-season residuals and a draw post-processor. The architecture that ran live until v0.5 launched. Picks engine: flat +10% edge across all bookie ranges.

Total picks: 3388
Settled: 3388
Hits: 1000
Hit rate: 29.5%
Total staked: £33880
Total PnL: −£1458.60
ROI: -4.31%

v0.5 (backtest)

Simulated

XGBoost + LogisticRegression hybrid, blended by live league-relative dominance (ppm_z gate). The LR component extrapolates beyond what trees can — fixes the runaway-leader compression on Bayern, PSG, etc. Tiered edge thresholds (+50% safe / +75% extreme) keep volume to a handful per week.

Total picks: 253
Settled: 253
Hits: 76
Hit rate: 30.0%
Total staked: £2530
Total PnL: +£326.60
ROI: +12.91%

v0.6 (backtest) — current production

Simulated

v0.5 hybrid + 10 nothing-to-play-for binary flags + a continuous season_phase axis. Encodes end-of-season motivation: title-locked, relegated, auto-promoted, playoff-locked, safe-no-climb. Flags fire on ~5% of fixtures (concentrated last 5–8 matchdays) and shift probability mass toward better-targeted picks-engine selections.

Backtest: these picks weren't placed in real time. They show what v0.6 would have selected on the 2024+2025 walk-forward fixtures using B365 closing odds. Treat as a calibration check, not banked profit.

Total picks: 258
Settled: 258
Hits: 80
Hit rate: 31.0%
Total staked: £2580
Total PnL: +£477.30
ROI: +18.50%

By pick category — v0.6 backtest

Category	Picks	Settled	Hits	Hit rate	Avg edge	PnL	ROI
favourite	0	0	0	—	—	+£0.00	—
reasonable	113	113	45	39.8%	+87.2%	+£192.50	+17.04%
longshot	145	145	35	24.1%	+121.0%	+£284.80	+19.64%

By league — v0.6 backtest

League	Picks	Hits	Hit rate	PnL	ROI
BL1	19	2	10.5%	−£114.50	-60.26%
EL1	39	10	25.6%	−£37.00	-9.49%
EL2	41	17	41.5%	+£299.30	+73.00%
ELC	55	13	23.6%	+£5.70	+1.04%
FL1	20	7	35.0%	+£81.50	+40.75%
PD	33	13	39.4%	+£108.40	+32.85%
PL	27	12	44.4%	+£131.90	+48.85%
SA	24	6	25.0%	+£2.00	+0.83%

How challenger versions work

A new model is trained on a richer feature set or a different architecture.
It runs through walk-forward backtest on the same historical fixtures the live model handled.
If it beats the live version on Brier score and per-cohort accuracy, it gets promoted to the next version number.
If it doesn't, it stays on the bench. Version numbers only go up when the math earns it.

We're in beta — this whole site is a public lab notebook for the picks engine. v0.4 and v0.5 history stays visible. v0.6 is now the live model and will accumulate fresh picks from here. Future versions (0.7+) will line up here for comparison whenever they show up.