Accuracy

Our model is retrained every gameweek. Once each gameweek concludes, every prediction is re-audited against the points players actually scored and benchmarked against FPL's official projection and the simple ratings managers use most, using industry-standard accuracy measures so the results can be compared directly.

How the model compares

Every week our predictions are scored next to the ways a manager can rate players without us: FPL's own projection, recent form and points per match. All are scored the same way, on the same players, from what was known before the deadline.

Us against FPL's projectionAhead on 3 of 4headline measures, over 1 gameweek
Us against recent formAhead on 4 of 4headline measures, over 9 gameweeks
Us against points per matchAhead on 4 of 4headline measures, over 9 gameweeks

Top pick6.9 ptsWhat the player we rated highest went on to score, on average each week.

FPL's projection 9.0 pts (us 8.0 pts)ahead 0 of 1 wk
Recent form 3.2 pts✓ ahead 5 of 9 wks
Points per match 1.0 pts✓ ahead 8 of 9 wks

Top 15 picks4.7 ptsAverage points of the 15 players we rated highest. The average player with a fixture scored 1.4.

FPL's projection 5.5 pts (us 6.3 pts)✓ ahead 1 of 1 wk
Recent form 4.0 pts✓ ahead 6 of 9 wks
Points per match 3.7 pts✓ ahead 7 of 9 wks

Right call, head to head69%Take any two players: how often the one we rated higher went on to score more. A coin toss gets 50%.

FPL's projection 71% (us 76%)✓ ahead 1 of 1 wk
Recent form 66%✓ ahead 8 of 9 wks
Points per match 64%✓ ahead 9 of 9 wks

Average miss2.2 ptsHow far a typical prediction was from the points the player actually scored. Lower is better.

FPL's projection 2.2 pts (us 2.0 pts)✓ ahead 1 of 1 wk
Recent form 2.6 pts✓ ahead 9 of 9 wks
Points per match 2.7 pts✓ ahead 9 of 9 wks

WeekTop 15 picks, points eachRight call, head to headAverage miss (shorter is better)

GW2backtest
Us
6.7 pts ✓
Form
3.8 pts
PPM
3.8 pts
Us
70% ✓
Form
67%

0 of 9 weeks are the model's live predictions, stored before the deadline. Weeks marked backtest were rebuilt afterwards from the match data available before the deadline, with the injury flags of the time where we hold a record of them and every player treated as fit where we do not, so injured players count against those weeks. FPL's projection is on record for 1 week; every comparison with it uses only those weeks.

All the figures, for comparing with other sources

Week	Source	MAE	RMSE	Pairwise	Top pick	Top-15 pts	Top-15	NDCG@15	Spearman
GW2	Us	2.315	3.280	0.704	13	6.67	0.33	0.605	0.51
	Recent form	3.082	4.738	0.671	0	3.80	0.07	0.222	0.44
	Points per match	3.082	4.738	0.671	0	3.80	0.07	0.222	0.44
GW3	Us	2.157	2.896	0.675	2

What the advice would have scored

For every team audited before a deadline: the captain and best XI we suggested against what the manager actually picked, scored with the points the players went on to get, and the first transfer we suggested, net of its hit. Lineups are compared before automatic substitutions.

The record starts once a gameweek with audits has finished.

What the numbers mean

Each measure in plain words, alongside its standard statistical name.

Top pick: Each week, the single player a source rated highest, and the points he actually scored, averaged over the weeks. The closest thing to asking who should have worn the armband.
Top 15 picks: The 15 players a source rated highest among everyone with a fixture, and the average points they actually scored. Set against the average player with a fixture, it shows how much a shortlist from that source was worth.
Right call, head to headpairwise accuracy, Kendall concordance: Take every pair of players owned by at least 1% of managers who finished on different points. How often did the source rate the higher scorer above the other? A coin toss scores 50%; ties in the rating count as half right.
Average missMAE, mean absolute error: For every player, take the gap between the points predicted and the points he actually scored, ignore whether it was too high or too low, and average those gaps. An average miss of 2.3 means a typical prediction was 2.3 points out. Lower is better. Only players owned by at least 1% of managers count, so empty benches don't flatter the number.
FPL's projectionep_next: The expected points the official game shows for each player. The last value published before the deadline is scored exactly like the model's, so the two can be compared on the same weeks.
Recent form: A player's average points per gameweek over the last 4 gameweeks, blanks included: the number most managers glance at first.
Points per match: A player's season points divided by the matches he has played in, as the official game shows it.
Live and backtest: A live week is one where the model's predictions were stored before the deadline. A backtest week is one the model never ran for at the time, rebuilt afterwards from the data that existed before the deadline. Where we hold the injury flags from before that deadline they are applied; otherwise every player is treated as fit, so injured players count against the model.
Big-miss errorRMSE, root mean square error: The same gaps as the average miss, but each is squared before averaging and the square root taken at the end, so one 10-point miss counts far more than five 2-point misses. Shown in the full figures.
Top-15 overlap and ranking qualityTop-15, NDCG@15: How many of a source's top 15 really finished in the week's top 15, and how much of the real top 15's points they captured with the very top counting most. Both swing a lot from week to week, so they are shown in the full figures.
Rank orderSpearman correlation: How closely a source's order of every owned player matches the order they really finished in, from −1 (exactly backwards) through 0 (no relation) to 1 (identical). Shown in the full figures.