Methods and calibration

1. Solver

The core optimiser is a rolling-horizon mixed-integer linear programme over an H-week window (default H = 5). The objective maximises the discounted sum of starting XI expected points less transfer-hit penalties.

maximise
  Σ_{t=1..H} γ^(t-1) * [ Σ_{i ∈ XI_t} ( E[P_{i,t}] * (1 + 1_{i = captain_t}) ) - 4.0 * max(0, Transfers_t - FreeTransfers_t) ]

subject to
  1. Roster size: Σ x_{i,t} = 15
  2. Budget: Σ cost_{i,t} * x_{i,t} + bank_t ≤ Budget_t
  3. Club limit: Σ_{i ∈ Club_k} x_{i,t} ≤ 3
  4. Formation: 1 GKP, 3–5 DEF, 2–5 MID, 1–3 FWD
  5. One captain, who must start

2. Live freeze versus reconstructive replay

Live pages lock odds, minutes and the plan two hours before the official deadline, hash the JSON, and do not edit after results. The 2025/26 pages use a reconstructed cutoff (first kickoff minus 90 minutes) built after the season. Fixtures come from the final export; private bank, exact pre-deadline prices and timestamped odds were not archived. See reconstructive replay and freeze and hash.

3. 2025/26 replay error

Computed from this reconstructive replay (GW2–GW38). GW1 had no locked expected-points objective. Player errors compare unmultiplied starting-XI points to the locked forecast.

Squad weeks

37

Squad MAE

17.2 pts

Squad bias

-15.0

Player-round MAE

3.25 (403 XI rows)

4. Illustrative position table

The table below is from the illustrative sample pack (5,520 labelled player-rounds). It is not computed from the 2025/26 ingest.