Four approaches. One season of evidence.

Prospective results from frozen weekly forecasts.

These are virtual experimental squads, separate from our two existing published teams. Every approach starts with the same 15 players and £100m budget. We freeze all four forecasts together, then track their decisions and outcomes.

Graded gameweeks: 4. Frozen weeks awaiting grades: None. Weeks not enrolled, recorded as dated gaps: GW5 (owner away 19 to 30 September 2026; no rehearsal or freeze was run for GW5).

Frozen forecast receipts

• GW4: e59f4913c3c2e7c0def8d477518946b758b85cefe3ddab08d2bd684b267b2d8c — captured 2026-09-11T12:31:49.966516+00:00

These hashes identify retained forecasts. Local timestamps alone do not prove when a forecast was published; independent publication must precede the deadline.

Lower forecast error is better; more squad points is better. Forecast losses are averaged with equal weight per gameweek on identical player populations. A lucky captain can win a week without proving that the underlying forecast is better. We do not automatically select a winner.

What the comparison can tell us

We report all players, players who averaged at least 60 minutes per completed gameweek before the deadline, and the union of the four selected squads. Cohorts use frozen inputs, never the minutes a player eventually played. The 60-minute cohort uses a gameweek average, not a per-match average; empty cohorts have no score.

Paired forecast differences

Each difference compares the same players in the same weeks. Negative mean squared error differences favor the first approach. These are exploratory comparisons, not a declaration of a winning model.

Equal gameweek weights. Exploratory 95% moving-block intervals use 3-week blocks; require 12 consecutive eligible weeks with unchanged model versions. Intervals are not adjusted for multiple comparisons and do not establish next-season performance. Unknown external model revisions suppress intervals.

Minutes research: candidate, not production

Trained on 2023/24 and evaluated on 2024/25 player-only single-fixture gameweeks. The transition model predicts no appearance, under 60 minutes, or 60+ minutes. These are played-minute roles, not start probabilities.

Lower is better. The candidate improves probability scores and RMSE but worsens MAE against the last-four baseline. Cold starts remain weak; injuries, blanks and doubles are not validated. It has not replaced our production minutes assumptions.

Fixture research: small historical improvement

A separate candidate adds home/away and lagged team/opponent goal rates. Each fixture in a double gets its own forecast, using only history available before the gameweek cutoff. Both models were trained on 2023/24; the same specification was evaluated on the later season.

The comparison isolates fixture features within a per-fixture model, not performance against our production gameweek model or odds. These are retrospective extracts with an estimated deadline cutoff. Historical blank weeks, cold starts and availability remain unvalidated. The candidate is research only; live forecasts are unchanged.

Live fixture shadow

Rehearsal only — not frozen. GW4, observed 2026-09-07T19:46:45.237862+00:00. Receipt-backed histories cover 654 players; 654 qualify for the fixture-model comparison. Cold-start fallbacks: 0. Blanks: 0.

This separate shadow compares fixture forecasts with the production statistical model on the same official inputs. It does not change the four squads. Cold starts copy production and are excluded from the fixture comparison. Forecast differences are not measured accuracy gains.

Snapshot: f39a4bc0cfe6a78f35edb026d456911a4018c0b6e69464b7fcc6cf8a91cbccd3. Raw histories and forecasts are retained privately. A local hash alone does not prove pre-deadline publication. This status describes the dated snapshot, not a continuously refreshed feed.

Lineup decision rehearsal

A scenario optimizer evaluates all 3,300 legal XI and outfield bench-order combinations, plus captain/vice pairs. It accounts for formation-constrained automatic substitutions and vice-captain fallback. This GW4 rehearsal uses assumed appearances and a separate random-seed audit.

These are simulated expectations, not observed gains. Decision changes require prospective validation. Shared-club dependence is a stress test, not an estimated correlation. The controlled four-arm policy remains unchanged.

Joint engine sample check

We now retain aligned appearance and points from the market engine itself. The retained samples reproduce its forecast means. The lineup optimizer uses 5000 draws and is audited on a separate seed with the same number of draws.

Candidate minus baseline: -0.018 simulated points. The approximate Monte Carlo interval is [-0.078, +0.042]. This measures sampling precision inside this simulator, not uncertainty about real-world gains.

These samples belong to the market model; they are not distributions for the statistical or external forecasts. Engine assumptions still require calibration. The four-arm policy is unchanged.

The controlled decision policy

Version 1 considers at most one positive-net transfer per week, then selects a legal XI, captain and vice using expected points. Bank balances, purchase and selling prices, free transfers and hits carry forward. Chips are disabled for every approach. This is a controlled comparison, not a complete season optimizer.

Model versions and any human intervention are recorded before the deadline. Existing history is never backfilled into this experiment. Missing providers, partial results, changed rules or missing weeks stop the pipeline rather than silently changing the comparison.

What the four forecasts use

The market arm combines match odds with official player statistics, with editorial overrides disabled. The statistical arm uses lagged points, minutes, xG and xA with coefficients fitted on 2023/24. Neither internal model uses FC27 ratings.

The hybrid is a prespecified 50/50 average, not an optimized blend. The reference averages official FPL projections with Fantasy Football IQ; players absent from FFIQ retain the official projection, with coverage recorded in the evidence. External providers' training and underlying data lineage are undisclosed. No approach has yet demonstrated a prospective advantage in this experiment.

Download the experiment summary · Published comparisons · Existing track record