Calculations

The calculations layer turns the typed sources.* tables into the analytical primitives that the marts ultimately consume. It splits into two families:

inplay_* — per match-minute

Model	Grain
inplay_1_by_minute	one row per (event, minute), forward/backward filled
inplay_2_score	one row per (event, minute) — derived score winner

matches_* — per event

Sequential pipeline; each step adds derived columns on top of the previous one.

#	Model	Adds
1	matches_1_odds_prematch	first/last prematch odds
2–4	matches_2_odds_inplay_home, matches_3_odds_inplay_draw, matches_4_odds_inplay_away	per-side in-play odds extremes + window stats
5	matches_5_score	parsed scores + winner side
6	matches_6_team_stats_window	rolling 1-year team scoring/conceding averages
7	matches_7_league_stats_window	rolling 1-year league averages
8	matches_8_team_strength	attack/defence strength relative to league
9	matches_9_expected_goals	expected goals per side & total
10	matches_10_overperformance	actual minus expected goals
11	matches_11_probability	implied probabilities + bookmaker margin
12	matches_12_prob_win	Poisson-derived match-outcome probabilities
13	matches_13_prob_goal_diff	Skellam goal-difference distribution
14	matches_14_prob_goals_total	Poisson total-goals distribution
15	matches_15_strat_expected_value	EV per outcome
16	matches_16_strat_kelly_criterion	Kelly fraction per outcome
17	matches_17_strat_predicted	predicted winner + meta
18	matches_18_strat_bet	bet sizing, P/L and edge

All models materialise as ClickHouse ReplacingMergeTree tables, partitioned by event_partition_key, built incrementally.