Model

All leagues Premier League LaLiga Bundesliga

Target: home possession %. Possession is zero-sum inside a match, so this one number defines the whole result — the away side is 100 minus it. This is a regression problem, not a win/draw/lose one. No model is trained yet; the figures below are what the modelling layer has to work with.

Training rows

Usable features

Fit split

Holdout split

Fixtures to score

What a model must beat — mean absolute error, percentage points

Always predict 50% The know-nothing floor		10.89
Home side's own recent average Ignores who they are playing		8.52
Naive midpoint Home form vs what the away side concedes — the one that matters		7.09

A model is only earning its keep below 7.09 points of error. Beating "always 50" is not an achievement.

Training set

Rows Played matches where both sides had enough prior matches to form an average	30
Home possession seen	25.9% – 73.1%
Mean / spread	51.4% ± 12.7
Earliest match	22 Aug 2026
Latest match	13 Sep 2026
Split Chronological, never random	24 / 6

68 features against 30 rows. More columns than examples means any model will fit the training data perfectly and learn nothing that generalises. Until the season fills out, prefer a handful of possession columns and strong regularisation over the full set — or widen training.season in config.yaml.

Feature inventory — 68 columns present in both training and fixtures

Minutes spent ahead, level and behind are deliberately not here. They are only known after the whistle, so using them to predict possession would be leakage. They are stored for explaining errors instead: a match settled early is inherently less predictable, because the leading side stops trying to keep the ball.

Possession (30)

home_poss_l5away_poss_l5home_poss_l10away_poss_l10home_poss_l20away_poss_l20home_poss_home_l5away_poss_away_l5home_pass_acc_l5away_pass_acc_l5home_passes_l5away_passes_l5home_touches_l5away_touches_l5home_long_balls_l5away_long_balls_l5home_poss_lost_l5away_poss_lost_l5home_dispossessed_l5away_dispossessed_l5home_recoveries_l5away_recoveries_l5home_poss_won_mid_l5away_poss_won_mid_l5home_passes_f3_l5away_passes_f3_l5poss_gap_l5poss_gap_l10poss_naive_l5h2h_home_poss_avg

Attacking form (7)

home_xg_l5home_xgd_l5home_xg_l10away_xg_l5away_xgd_l5away_xg_l10h2h_home_xg_avg

Defensive form (4)

home_xga_l5home_xga_l10away_xga_l5away_xga_l10

Points and results (6)

home_ppg_l5home_ppg_l10away_ppg_l5away_ppg_l10ppg_edge_l5h2h_home_ppg

Volume (6)

home_shots_l5home_sot_l5home_corners_l5away_shots_l5away_sot_l5away_corners_l5

Matchup edges (3)

xgd_edge_l5home_attack_vs_away_defenceaway_attack_vs_home_defence

Head to head (2)

h2h_meetingsh2h_home_goals_avg

Schedule and fatigue (6)

home_days_restaway_days_resthome_matches_last_14daway_matches_last_14dhome_prev_europeanaway_prev_european

Other (4)

home_matches_beforehome_finishing_gapaway_matches_beforeaway_finishing_gap

Stake guidance — open lines, rated out of 10

Two separate things have to be true before staking much: the edge must be large, and the model must have shown it can find real edges. The first is arithmetic; the second is a track record. The rating is the smaller of the two, so a big edge from a model with no record scores low — correctly. It is 14 settled bets in, so the ceiling is low for a while yet.

Kickoff	Match	Line is about	Line	Under	Over	Model	Market	Bet	Edge	Stake	Why
14 Sep 20:00	LEE v NEW	Newcastle United	52.5	1.70	2.01	57.8%	46% over	OVER	+25.0%	2/10	edge alone suggests 5/10, capped at 2 by a very thin track record (14 settled)
13 Sep 16:30	MUN v MCI	Manchester City	56.5	2.01	1.70	58.6%	54% over	OVER	+1.6%	1/10	edge of +1.6% at 1.70

0/10 means the price is against you — not a close call, a bet to skip. Anything above that is a fraction of what you would stake at full confidence, and quarter-Kelly is treated as the ceiling even at 10/10, because Kelly assumes the probability is exactly right and this one is an estimate.

Next fixtures — with the naive prediction to beat

Kickoff	Match	Home form	Away form	Gap	Naive home poss
13 Sep, 16:30	Manchester United v Manchester City	56.5%	64.5%	-7.9	46.0%
14 Sep, 20:00	Leeds United v Newcastle United	42.9%	49.9%	-7.0	46.5%
18 Sep, 20:00	Brentford v Chelsea	45.1%	46.8%	-1.7	49.2%
19 Sep, 12:30	Tottenham Hotspur v Aston Villa	57.8%	48.2%	+9.6	54.8%
19 Sep, 15:00	Brighton and Hove Albion v Arsenal	67.2%	59.4%	+7.8	53.9%
19 Sep, 15:00	Newcastle United v Hull City	49.9%	29.4%	+20.5	60.2%
19 Sep, 15:00	Everton v Ipswich Town	43.2%	44.0%	-0.8	49.6%
19 Sep, 17:30	Nottingham Forest v Coventry City	43.9%	39.4%	+4.5	52.3%
20 Sep, 14:00	Manchester City v Sunderland	64.5%	48.2%	+16.3	58.1%
20 Sep, 14:00	Leeds United v Crystal Palace	42.9%	41.5%	+1.4	50.7%