Predictive Signals

Some attempt to see if there are any predictive qualities to the stats we have available. Shared with everyone, because I'm not certain it's useful. We look at the "ICT Index" that everyone ignores in the main game, plus a rolling average of form, and the difficulty of the next fixture. And it seems there is some sort of predicitve element there. Or, it might all be dumb and circular - of course these things predict the points, because they're the things used to calculate the points... Whatever, I just asked the Gimp to create it for me, so 🤷‍♀️

Team Form Pulse Check

This is the basic stuff - teams ranked by their players' average form. And the three most form-y players for each of them. And then a chart plotting that form vs. how hard their next five fixtures are.

Rank Team Active Players Team Form Top Form Players
1 Wolves 21 68.2
Toti DEF (8.0)
Edozie MID (6.0)
André MID (5.5)
2 Arsenal 19 68.1
Gabriel DEF (5.8)
Saka MID (5.8)
J.Timber DEF (5.5)
3 West Ham 20 67.8
Todibo DEF (9.0)
Diouf DEF (6.6)
Hermansen GKP (6.2)
4 Liverpool 19 64.2
Virgil DEF (7.2)
Mac Allister MID (5.8)
Frimpong DEF (5.5)
5 Bournemouth 19 61.4
Semenyo MID (7.0)
Hill DEF (6.4)
Truffert DEF (5.4)
6 Man City 19 58.9
O’Reilly DEF (9.5)
Haaland FWD (6.0)
AĂŻt-Nouri DEF (5.6)
7 Man Utd 19 58.0
B.Fernandes MID (7.0)
Šeško FWD (5.4)
Mbeumo MID (4.6)
8 Chelsea 22 57.3
JoĂŁo Pedro FWD (9.2)
Palmer MID (9.0)
James DEF (4.7)
9 Sunderland 22 55.6
Ellborg GKP (10.0)
O'Nien DEF (5.0)
Ballard DEF (4.4)
10 Crystal Palace 16 52.1
Henderson GKP (6.2)
Sarr MID (5.8)
Lacroix DEF (5.7)
11 Nott'm Forest 23 50.8
Anderson MID (5.4)
Gibbs-White MID (4.8)
Morato DEF (4.0)
12 Everton 16 50.7
Dewsbury-Hall MID (6.8)
McNeil MID (5.5)
Garner MID (4.8)
13 Brighton 22 50.3
Van Hecke DEF (5.0)
Gomez MID (4.6)
Welbeck FWD (4.4)
14 Brentford 18 48.8
Damsgaard MID (4.8)
O.Dango MID (4.6)
Van den Berg DEF (4.4)
15 Leeds 19 48.7
Okafor MID (6.5)
Stach MID (5.7)
Gruev MID (3.8)
16 Newcastle 19 46.8
Bruno G. MID (7.0)
Gordon MID (4.0)
Thiaw DEF (4.0)
17 Fulham 21 45.4
Iwobi MID (6.2)
Wilson MID (6.0)
RaĂşl FWD (4.0)
18 Aston Villa 20 42.0
Mings DEF (4.0)
Martinez GKP (3.8)
Rogers MID (3.4)
19 Burnley 18 36.5
Anthony MID (5.2)
Hannibal MID (5.0)
Flemming FWD (4.8)
20 Spurs 21 29.5
Gray MID (4.6)
P.M.Sarr MID (3.2)
Solanke FWD (2.8)

Coming good

Players whose 3-game average is outperforming their longer-term baseline.

  • J.Timber DEF ARS
    3G mean 8.0 vs 5G avg 5.4
    +2.6
  • Damsgaard MID BRE
    3G mean 7.33 vs 5G avg 4.8
    +2.53
  • Stach MID LEE
    3G mean 5.67 vs 5G avg 3.4
    +2.27
  • AndrĂ© MID WOL
    3G mean 6.67 vs 5G avg 4.4
    +2.27
  • Mac Allister MID LIV
    3G mean 8.0 vs 5G avg 5.8
    +2.2

Coming too soon

Players whose short-term form has slipped under their baseline.

  • Palmer MID CHE
    3G mean 4.33 vs 5G avg 9.0
    -4.67
  • Bruno G. MID NEW
    3G mean 0.0 vs 5G avg 2.8
    -2.8
  • Okafor MID LEE
    3G mean 0.0 vs 5G avg 2.6
    -2.6
  • Mings DEF AVL
    3G mean 0.67 vs 5G avg 3.2
    -2.53
  • Madueke MID ARS
    3G mean 0.33 vs 5G avg 2.8
    -2.47

Predictive signal leaderboard

Of the signals we're looking at, which players are doing well? All the input features, mulched together into an overall score.

# Player Pos Club Signal Score
1 JoĂŁo Pedro FWD CHE 43.27
2 Sarr MID CRY 25.74
3 Garner MID EVE 20.24
4 Anderson DEF NFO 19.79
5 Garnacho MID CHE 19.68
6 André MID WOL 19.61
7 B.Fernandes MID MUN 19.6
8 Rodrigo MID MCI 19.41
9 Wharton MID CRY 18.62
10 Palmer MID CHE 18.25
11 Gibbs-White MID NFO 18.2
12 M.Salah MID LIV 16.98
13 Tarkowski DEF EVE 16.16
14 Dewsbury-Hall MID EVE 16.16
15 Gordon MID NEW 16.13
16 Summerville MID WHU 16.07
17 Solanke FWD TOT 16.01
18 Alderete DEF SUN 14.68
19 J.Timber DEF ARS 14.57
20 Bernardo MID MCI 14.47

Projected next-match points

And the juice - can we use that score to predict how a player will do next match? It runs their current performance score against whether the next match is home/away, its diffculty, etc. and gives a points forecast.

# Player Pos Club Next Opponent Predicted Pts Range
1 João Pedro FWD CHE Home vs NEW 3 17.95 17.51 – 18.38
2 Sarr MID CRY Home vs LEE 2 14.55 14.11 – 14.98
3 Anderson DEF NFO Home vs FUL 2 11.59 11.16 – 12.03
4 André MID WOL Away vs BRE 4 10.92 10.48 – 11.36
5 Rodrigo MID MCI Away vs WHU 2 9.96 9.52 – 10.39
6 Wharton MID CRY Home vs LEE 2 9.68 9.25 – 10.12
7 Gibbs-White MID NFO Home vs FUL 2 9.36 8.93 – 9.8
8 Garner MID EVE Away vs ARS 5 9.17 8.73 – 9.6
9 Tarkowski DEF EVE Away vs ARS 5 8.98 8.54 – 9.41
10 Palmer MID CHE Home vs NEW 3 8.93 8.5 – 9.37
11 Gabriel DEF ARS Home vs EVE 3 8.42 7.98 – 8.86
12 Dewsbury-Hall MID EVE Away vs ARS 5 8.4 7.96 – 8.83
13 Summerville MID WHU Home vs MCI 4 8.39 7.96 – 8.83
14 J.Timber DEF ARS Home vs EVE 3 8.24 7.8 – 8.67
15 M.Salah MID LIV Home vs TOT 3 8.2 7.76 – 8.63
16 Casemiro MID MUN Home vs AVL 3 8.1 7.66 – 8.54
17 Henderson GKP CRY Home vs LEE 2 7.21 6.77 – 7.64
18 Ballard DEF SUN Home vs BHA 3 7.04 6.6 – 7.47
19 Saka MID ARS Home vs EVE 3 7.03 6.6 – 7.47
20 Solanke FWD TOT Away vs LIV 4 6.86 6.42 – 7.3

Model scorecard

So... how good is the model, if at all?

The MAE values shown in the scorecards are literal FPL points (they average how far the predictions miss by). During training we learn the link between a match's ICT signals and that same match's score; for future fixtures we reuse each player's latest snapshot and swap in the upcoming opponent's difficulty/home-or-away flag to make a forward-looking call.

Because most players only score a couple of points, we also slice the errors by actual and predicted score bands, track how often we flag genuine big hauls (≥8 pts), and check whether the model's top picks overlap with the actual top performers each gameweek.

Prediction scorecard

Back-tested on recent gameweeks. MAE is the average miss in FPL points; the hit-rate shows the share of predictions that landed within two points of reality.

GW MAE Hit rate Samples
25 0.97 65.0% 811
26 1.01 66.4% 896
27 0.96 64.8% 817
28 0.92 67.0% 818
29 1.09 63.9% 819

Largest misses help highlight outliers the model struggles with.

  • Lewis-Potter
    GW 17 vs WOL
    DEF · BRE
    Δ 18.54 pts
    2.46 → 21.0
  • Hudson-Odoi
    GW 16 vs TOT
    MID · NFO
    Δ 18.13 pts
    0.87 → 19.0
  • Schade
    GW 18 vs BOU
    MID · BRE
    Δ 17.85 pts
    2.15 → 20.0
  • JoĂŁo Pedro
    GW 29 vs AVL
    FWD · CHE
    Δ 17.83 pts
    1.17 → 19.0
  • Eze
    GW 13 vs CHE
    MID · ARS
    Δ 17.3 pts
    19.3 → 2.0

Big haul classification

How often the model correctly flags players expected to hit 8+ points.

≥ 8 pts

0.175

Precision (hauls we called correctly)

0.1

Recall (share of all hauls we spotted)

0.128

F1 (balance of precision & recall)

Predicted hauls: 280 · Actual hauls: 488 · True positives: 49

Scores run 0–1; higher is better. Precision/recall/F1 around 0.25–0.35 would be solid for noisy haul calls, while anything under ~0.1 means the model is mostly guessing.

Top-10 overlap

How many of the actual top scorers we catch in the model's top picks each gameweek.

Top 10

0.082

Avg recall (actual top scorers recovered)

0.082

Avg precision (how many picks really hauled)

Both metrics run 0–1; higher is better. A healthy shortlist would sit around 0.4–0.6 recall/precision, while numbers below ~0.2 suggest the picks aren’t much better than luck.

Error by actual score bucket

Each bar shows how far off the predictions were (average absolute error) for players who actually landed in each points band, plus how many fell inside that band. In other words, it checks whether the model stays sharp for low scorers as well as the rare big hauls. Does the miss size stay sensible no matter how many points the player truly scored?

Lower is better: MAE around 1–2 points in most buckets is respectable; when errors regularly creep above 3–4 points the model is missing the mark.

Calibration at higher predictions

Points are grouped by what the model predicted (e.g. 6–9 points) and compared with what actually happened, so we can see if confident forecasts sit too high, too low, or on target. It is a gut check on whether the model is overhyping or under-calling good outings. When we predict 6–9 points, do the real scores usually end up in that range?

A well-calibrated model clusters near the diagonal with average miss under ~2 points; if points sit far above or below the band lines, the model is overconfident or under-confident.

Top-10 overlap by gameweek

The two lines track, week by week, how many of the real top scorers appear in the model's top picks (recall) and how many of the model's picks actually went on to haul (precision). It reveals whether the shortlist consistently finds the right players or just gets lucky. Are the model's top choices reliably capturing the real stars each gameweek?

Good weeks hover near or above 0.5 on both lines; dips below ~0.2 hint the model’s weekly picks are more miss than hit.

Linear R²

0.80

Explains how much variance the straight-line model captures using the core signals.

Linear MAE

0.54

Average miss (in FPL points) for the baseline regression.

XGBoost R²

0.83

How much variance the tree ensemble explains once we allow nonlinear interactions.

XGBoost MAE

0.44

Average miss for the boosted model — lower means sharper projections.

Feature spotlight

And how much do all the features matter?

SHAP values show the average boost each feature gives across all predictions. They apportion credit fairly, even when features interact.

Influence 36.4% of impact
0.667
Rolling Mean Points 3 33.0% of impact
0.604
Ict Index 19.7% of impact
0.361
Creativity 3.9% of impact
0.072
Rolling Sum Points 5 2.7% of impact
0.049
Fixture Difficulty 2.2% of impact
0.04
Threat 2.0% of impact
0.036
Is Home 0.0% of impact
0.001
Trained on 22,233 player fixtures, the baseline linear model already explains 0.80 of the variance with a 0.54-point MAE, while XGBoost pushes accuracy to R² 0.83 and MAE 0.44. Influence currently drives the model's decisions. Rolling Mean Points 3 is the runner-up and adds extra context to the predictions.

Glossary

The signals use a mix of official FPL metrics and statistical jargon. Here's a quick refresher.

ICT Index
Fantasy Premier League's blend of Influence, Creativity, and Threat metrics to gauge how involved a player is in decisive actions.
Influence
Measures how heavily a player affects match outcomes (goals, assists, key contributions). High influence means the player drives team results.
Creativity
Tracks the rate of chance creation—crosses, key passes, and set-piece threat. Assisters tend to spike here.
Threat
Quantifies how likely a player is to score based on shots and positioning inside dangerous zones.
Fixture Difficulty
Club-provided rating (1 easiest — 5 hardest) estimating how tough the opponent is for that particular match.
Rolling Mean
Moving average across the last N games. A 3-game rolling mean smooths noisy weekly scores into a clearer form signal.
Rolling Sum
Moving total across the last N games. A 5-game sum captures medium-term consistency vs. short bursts.
SHAP
Shapley Additive exPlanations: a model-agnostic method that shows how much each feature pushed a prediction up or down.
MAE
Mean Absolute Error — the average absolute difference between predicted and actual points.
R²
Coefficient of determination. Shows how much of the variation in points the model explains (1.0 means perfect).
XGBoost
Extreme Gradient Boosting — an ensemble of shallow decision trees trained sequentially to reduce errors, great at spotting nonlinear patterns.