How it works · every forecast tested on past seasons · weak results included

How dugout knows what it says.

The rule

AI writes the words, never the numbers. Every figure is filled straight from the data, and nothing is published unless two independent checks pass.

Stat Auditor · no language model
Every cited number's query is re-run against its source and must reproduce the stored value; goals, minutes, availability and fees are compared with a second source; any digit or number word not filled from the ledger fails the build.
Fact-Check · DeBERTa-v3 NLI + BGE retrieval
Every other claim needs one official source or two independent ones (one vote per ownership group). Contradicted claims withhold the section; unsupported sentences are cut. False verifications on hand-labelled football claims: .

On held-out questions the system routed correctly, answered , refused every unanswerable question, and published untraceable numbers. The open-weight LLM's drafts passed every check for of sections; deterministic templates wrote the rest.

The models

Match predictor
Elo + expected-goals Dixon–Coles stack, refit every four gameweeks. Ranked probability score (lower is better):2024-25: dugout · Elo · closing odds 2025-26: dugout · Elo · closing odds
beats Elo · bookmakers still better
Forecasts in other leagues
Same replay, each league's own history (no expected goals outside the Premier League, so the stack and Elo come out almost level). Ranked probability score, last season:La Liga: stack · Elo · closing odds · shown: stackSerie A: stack · Elo · closing odds · shown: EloBundesliga: stack · Elo · closing odds · shown: EloLigue 1: stack · Elo · closing odds · shown: EloEredivisie: stack · Elo · closing odds · shown: Elo
the stack where it wins · Elo elsewhere
Valuation
Gradient boosting on past fees. Median error on test transfers vs for “fee = market value”. Every price comes with a likely range, wider for cheap players, whose fees vary most: on test transfers of fees landed inside it (aim: ).
beats market value
Transfer realism
P(sold in a window) × the club's share of plausible buyers. Sale model AUC ; destination AUC .
calibrated parts
Rumour brief
Every rumour on a club page comes with its source's label, how many independent outlets (one per media group) carry the link and since when, our estimate of the move (the transfer realism above), his contract, how often his club sells, and what became of the source's earlier reports on the same club. A source is rated against what similar reports did: on the latest season that estimate scored a Brier of vs for the recent base rate; AUC . Reliable or speculative needs a record clearly apart from that, beyond luck; almost every outlet, and the column's own round-ups, come out about average. The labels by outlet.
beats the base rate · few sources stand out
Player potential
Projected market value. Median error two years ahead (age curve ); four years .
beats age curve · wide bands
Playing-time outlook
Expected minutes share at a new club. R² vs for last season's minutes alone.
the validated part of fit
Style similarity
Manager fingerprint from Opta-derived team shares. It does not predict minutes (Spearman ), so it is shown as description, never as a forecast.
descriptive only
Injury risk
Load and history model. AUC , top-decile lift ; shown as risk tiers because calibration is limited. Other leagues have no current minutes, so their tiers come from the injury record and age alone: La Liga AUC , Serie A AUC , Bundesliga AUC , Ligue 1 AUC ; not rated in Eredivisie (too little injury data).
tiers, not percentages
Lineup optimiser
CP-SAT integer programme. Predicted XI overlap with the real XI vs for recent minutes.
slightly beats heuristic
Predicted XI in other leagues
Without FPL minutes, the best rule found is the club's last league XI (recency-weighted starts did not beat it). Share of the real XI predicted last season ; players API-Football lists as missing are replaced by the most recent other starter.
the last XI, absentees replaced
Player quality (VAEP)
Action values on open event data. Scoring-model AUC .
beats base rate
Roles & similarity
Gaussian mixtures on per-90 profiles. Role stability (adjusted Rand) .
stable roles

Full reports, including what did not work (FootballBERT embeddings were not adopted: the signal was leakage): docs/backtests

How dugout knows what it says.

The rule

The models

Match predictor

Forecasts in other leagues

Valuation

Transfer realism

Rumour brief

Player potential

Playing-time outlook

Style similarity

Injury risk

Lineup optimiser

Predicted XI in other leagues

Player quality (VAEP)

Roles & similarity