Calculations

Five weighted sub-metrics result in the overall score, known as the CompositeScore, as one number:

• Base difficulty — the difficulty metric from FPL directly, to use as a base. A fetch from api/fpl.js reaches fantasy.premierleague.com/api/fixtures through Vercel, and harvests the team_h_difficulty and team_a_difficulty. This is then converted to a 10 if 1, 30 if 2, 50 if 3, 70 if 4, and 90 if 5.
• Counter-matchup — this is the heaviest calculation. What this does, rather than defining a team's strength, it instead calculates where a team is strongest and weakest. For example, if a team who concedes mostly set pieces plays against a team that scores mostly from set pieces, the counter-matchup score will swing strongly one way, but if two teams' strengths and weaknesses don't counter each other, then this will remain balanced. The relevant data is fetched from understat.com's getTeamData endpoint. It will fetch shots, goals, situation, and total predicted goals from those shots, and calculate the average predicted goals for that situation, known as shared xG — otherwise where the biggest threat comes from for that team. It will then do this for all situations such as transition play (fast play) and box occupation (close up slow play) as well as set pieces, both for attacking and conceding, calculate a score out of 100 based on how good or bad the counter is, and then calculate the average of all the counters, resulting in Counter-matchup. This also has a maturity value of 10 fixtures, elevating to a max 0.25 weight towards the total CompositeScore in a linear ascent, to weaken the contribution of wild early-season data.
• Team form — recent trajectory, opponent-adjusted. It exists to correct base difficulty's preseason-seeded priors in-season. It's simply fetched from fantasy.premierleague.com/api through Vercel, pulls the bootstrap-static/ endpoint, and harvests the strength_overall_home and strength_overall_away values. It then runs oppAdj = oppStrength / leagueAvgStrength in order to swing form one way or the other based on whether the form is against stronger teams or weaker teams. This also has a maturity, but of 5 fixtures, to reach a max weight of 0.2 — to account for teams that appear in perfect form after only 1 or 2 games.
• Head-to-head — real cross-season meetings from the Understat history window. Kept deliberately small, as testing revealed that this weakens prediction accuracy if this holds more than a weight of 0.1. The last 10 meetings (or last 5 seasons) are fetched from understat.com through their getLeagueData/EPL/{season} endpoint, looping until 5 seasons are fetched. The head-to-head score is then calculated by 100 × (TotalPointsWon / TotalPossiblePoints), to account for the validity of all wins, losses, and draws.
• Home and away — home advantage strength compares how many points a team wins per game at home versus away, then measures how big that gap is compared to every other team in the league. A bigger gap means venue matters more for that team. Both teams' gaps get averaged and turned into a boost for the home side and an equal penalty for the away side, centred around a neutral 50.

A stacking penalty then applies across the sub-metrics. It and the base-difficulty weight are a matched pair — a weight that large would otherwise make a strong favourite's score nearly immovable, and the penalty is what lets several poor secondary signals tip a fixture without letting any single one do it alone. Neither should be tuned without re-checking the other.

The result maps to five bands, symmetric around a neutral midpoint, which drive colour everywhere in the app: great 75–100, good 60–74, neutral 41–59, tough 26–40, brutal 0–25. Thresholds are config rather than literals, so the palette can be recalibrated after a season's distribution has actually been observed.

There used to be a sixth signal, style clash, and removing it is the most instructive decision in the model. It substantially restated the counter-matchup — two of its three rules described the same interaction, read from the same source data, at half the weight — and two correlated weak predictors are worse than one, because the shared noise gets counted twice while the dependable metrics are diluted. It also assumed season-average pressing is a stable team property, which it is not: a side presses a promoted team and sits off against a title contender. And its observed spread was narrow enough that it could move the composite by roughly half a point — too small to change a decision, large enough to add noise whenever it read a fixture wrong.