How predictions work

The technical layer, without taking over the homepage.

FPL Copilot blends historical FPL scoring, recent player form, and fixture context to generate weekly player projections. This page gives a recruiter-friendly view of the model inputs, evaluation snapshot, and data freshness without turning the whole app into a technical dashboard.

Inputs

Recent FPL performance, fixture difficulty, ownership, price, minutes, and rolling player form all help shape the weekly predictions.

Model flow

The app ingests FPL data, refreshes its warehouse tables, trains the latest model, and stores scored predictions for the recommendation and optimizer flows.

Why it matters

Predictions are most useful for tie-breakers: captaincy calls, shortlists, differentials, and budget-efficient squad construction.

Model status

A compact view of prediction freshness and whether you are looking at live storage or a saved snapshot.

Prediction gameweek

Pending

Not available yet

Model version

Pending

Waiting for training metadata

Last ingestion

Pending

Not available yet

Serving mode

Pending

Live data path available.

Evaluation snapshot

Lower MAE and RMSE are better. These compare the baseline against the trained model on held-out gameweeks.

Baseline MAE

RMSE pending

Pending

Model MAE

RMSE pending

Pending

Validation gameweeks

Train rows: 0 ยท Validation rows: 0

Pending