File size: 8,944 Bytes
f40129e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
"""export_dataset.py - generates the portable, DERIVED/synthetic sample dataset.

This script does NOT read any live database, and it does NOT copy or
transform raw rows from football-data.co.uk or any other third-party
match-data provider. It generates a synthetic-but-realistic set of
engineered match features and 1X2 outcome
labels, distributed to match the empirical priors used in production
(class balance ~46% home win / 27% draw / 27% away win, which mirrors
publicly documented long-run baselines for the 5 European top-flight
leagues the production pipeline covers).

Why synthetic and not a real-data aggregate:
  - football-data.co.uk's commercial-use / redistribution terms for this
    project were flagged as unverified in the ingestion code itself
    (ml-service/app/ingestion/datasets/football_results_2018_2026.py,
    LICENSE = "Verificare commercial use con football-data.co.uk").
  - The other live-odds third-party source integrated elsewhere in the
    monorepo pipeline is explicitly non-redistributable in any form (its
    ingestion config flags it `is_public_redistribution_allowed=False`)
    and no row from it is ever touched by this export script.
  - Production Postgres (sport_intelligence.fixture) is never read from a
    public repo context.
  - Generating a synthetic sample sidesteps the licensing question
    entirely: no third-party row, raw or aggregated, is shipped. Only the
    feature SCHEMA (column names/semantics) is derived from the real
    pipeline (app/sport_intelligence/features/db_extractor.py
    FEATURE_NAMES_V2), which is this project's own original engineering
    work, not third-party data.

The 25 features generated here match FEATURE_NAMES_V2 from
app/sport_intelligence/features/db_extractor.py:
  dc_p_home, dc_p_draw, dc_p_away, elo_p_home, elo_p_draw, elo_p_away,
  imp_home, imp_draw, imp_away, home_form5_pts, away_form5_pts,
  home_form5_goals_for, away_form5_goals_for, home_form5_goals_against,
  away_form5_goals_against, home_form5_avg_xg, away_form5_avg_xg,
  home_form5_avg_xga, away_form5_avg_xga, home_days_rest, away_days_rest,
  h2h_home_win_rate_5, h2h_avg_total_goals_5, home_lineup_rating,
  away_lineup_rating

Usage:
    python export_dataset.py --rows 4000 --seed 42 --output data/derived_sample.csv
"""

from __future__ import annotations

import argparse
import csv
import random
from pathlib import Path

FEATURE_NAMES_V2 = [
    "dc_p_home", "dc_p_draw", "dc_p_away",
    "elo_p_home", "elo_p_draw", "elo_p_away",
    "imp_home", "imp_draw", "imp_away",
    "home_form5_pts", "away_form5_pts",
    "home_form5_goals_for", "away_form5_goals_for",
    "home_form5_goals_against", "away_form5_goals_against",
    "home_form5_avg_xg", "away_form5_avg_xg",
    "home_form5_avg_xga", "away_form5_avg_xga",
    "home_days_rest", "away_days_rest",
    "h2h_home_win_rate_5", "h2h_avg_total_goals_5",
    "home_lineup_rating", "away_lineup_rating",
]

# Empirical long-run class prior for the 5 covered leagues (documented in
# DATA_PROVENANCE.md). NOT derived from a specific third-party file - used
# only to shape the synthetic label distribution realistically.
_HOME_WIN_PRIOR = 0.46
_DRAW_PRIOR = 0.27
# away win = remainder


def _sample_team_strength(rng: random.Random) -> float:
    """Latent team strength on an Elo-like scale (mean 1500, sd 120)."""
    return rng.gauss(1500.0, 120.0)


def _softmax3(a: float, b: float, c: float) -> tuple[float, float, float]:
    import math

    m = max(a, b, c)
    ea, eb, ec = math.exp(a - m), math.exp(b - m), math.exp(c - m)
    s = ea + eb + ec
    return ea / s, eb / s, ec / s


def generate_row(rng: random.Random) -> dict[str, float | int | str]:
    """Generate one synthetic match row: 25 features + outcome label."""
    home_strength = _sample_team_strength(rng)
    away_strength = _sample_team_strength(rng)
    home_advantage = 65.0

    diff = (home_strength + home_advantage) - away_strength

    # Dixon-Coles-like probability triple derived from the latent diff,
    # softened toward the empirical draw prior (mirrors the production
    # dixon_coles.py + training.py EloTable.win_probability soft-draw
    # carving logic, re-implemented independently here for synthetic data).
    p_draw = max(0.08, _DRAW_PRIOR - 0.0003 * abs(diff))
    remaining = 1.0 - p_draw
    p_home_raw = 1.0 / (1.0 + 10 ** (-diff / 400))
    p_home = p_home_raw * remaining
    p_away = remaining - p_home

    # Elo-flavoured triple: same latent diff, independent noise draw so it
    # is correlated with but not identical to the Dixon-Coles triple
    # (mirrors two independently-fit sub-models feeding one meta-learner).
    elo_noise = rng.gauss(0.0, 15.0)
    elo_diff = diff + elo_noise
    elo_p_draw = max(0.08, _DRAW_PRIOR - 0.0003 * abs(elo_diff))
    elo_remaining = 1.0 - elo_p_draw
    elo_p_home_raw = 1.0 / (1.0 + 10 ** (-elo_diff / 400))
    elo_p_home = elo_p_home_raw * elo_remaining
    elo_p_away = elo_remaining - elo_p_home

    # Market-implied triple: a devigged noisy blend of the two above,
    # standing in for bookmaker consensus.
    imp_home, imp_draw, imp_away = _softmax3(
        (p_home + elo_p_home) / 2 + rng.gauss(0.0, 0.03),
        (p_draw + elo_p_draw) / 2 + rng.gauss(0.0, 0.03),
        (p_away + elo_p_away) / 2 + rng.gauss(0.0, 0.03),
    )

    # Rolling-form features, loosely correlated with latent strength.
    home_form5_pts = max(0.0, min(15.0, 1.5 * 5 + (home_strength - 1500) / 60 + rng.gauss(0, 2.0)))
    away_form5_pts = max(0.0, min(15.0, 1.5 * 5 + (away_strength - 1500) / 60 + rng.gauss(0, 2.0)))
    home_form5_goals_for = max(0.0, 6.0 + (home_strength - 1500) / 100 + rng.gauss(0, 1.5))
    away_form5_goals_for = max(0.0, 6.0 + (away_strength - 1500) / 100 + rng.gauss(0, 1.5))
    home_form5_goals_against = max(0.0, 6.0 - (home_strength - 1500) / 100 + rng.gauss(0, 1.5))
    away_form5_goals_against = max(0.0, 6.0 - (away_strength - 1500) / 100 + rng.gauss(0, 1.5))
    home_form5_avg_xg = max(0.0, 1.3 + (home_strength - 1500) / 300 + rng.gauss(0, 0.25))
    away_form5_avg_xg = max(0.0, 1.3 + (away_strength - 1500) / 300 + rng.gauss(0, 0.25))
    home_form5_avg_xga = max(0.0, 1.3 - (home_strength - 1500) / 300 + rng.gauss(0, 0.25))
    away_form5_avg_xga = max(0.0, 1.3 - (away_strength - 1500) / 300 + rng.gauss(0, 0.25))

    home_days_rest = max(2.0, rng.gauss(7.0, 2.0))
    away_days_rest = max(2.0, rng.gauss(7.0, 2.0))

    h2h_home_win_rate_5 = min(1.0, max(0.0, 0.5 + (diff / 800) + rng.gauss(0, 0.1)))
    h2h_avg_total_goals_5 = max(0.0, 2.5 + rng.gauss(0, 0.6))

    home_lineup_rating = home_strength / 1500.0 - 1.0 + rng.gauss(0, 0.05)
    away_lineup_rating = away_strength / 1500.0 - 1.0 + rng.gauss(0, 0.05)

    # Draw outcome from the Dixon-Coles-like triple (ground truth label).
    u = rng.random()
    if u < p_home:
        outcome = 0  # home
    elif u < p_home + p_draw:
        outcome = 1  # draw
    else:
        outcome = 2  # away

    return {
        "dc_p_home": p_home, "dc_p_draw": p_draw, "dc_p_away": p_away,
        "elo_p_home": elo_p_home, "elo_p_draw": elo_p_draw, "elo_p_away": elo_p_away,
        "imp_home": imp_home, "imp_draw": imp_draw, "imp_away": imp_away,
        "home_form5_pts": home_form5_pts, "away_form5_pts": away_form5_pts,
        "home_form5_goals_for": home_form5_goals_for, "away_form5_goals_for": away_form5_goals_for,
        "home_form5_goals_against": home_form5_goals_against, "away_form5_goals_against": away_form5_goals_against,
        "home_form5_avg_xg": home_form5_avg_xg, "away_form5_avg_xg": away_form5_avg_xg,
        "home_form5_avg_xga": home_form5_avg_xga, "away_form5_avg_xga": away_form5_avg_xga,
        "home_days_rest": home_days_rest, "away_days_rest": away_days_rest,
        "h2h_home_win_rate_5": h2h_home_win_rate_5, "h2h_avg_total_goals_5": h2h_avg_total_goals_5,
        "home_lineup_rating": home_lineup_rating, "away_lineup_rating": away_lineup_rating,
        "outcome": outcome,
    }


def export_dataset(rows: int, seed: int, output: Path) -> None:
    rng = random.Random(seed)
    output.parent.mkdir(parents=True, exist_ok=True)
    fieldnames = FEATURE_NAMES_V2 + ["outcome"]
    with output.open("w", newline="", encoding="utf-8") as f:
        writer = csv.DictWriter(f, fieldnames=fieldnames)
        writer.writeheader()
        for _ in range(rows):
            writer.writerow(generate_row(rng))
    print(f"Wrote {rows} synthetic rows to {output}")


def _parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--rows", type=int, default=4000)
    parser.add_argument("--seed", type=int, default=42)
    parser.add_argument("--output", type=Path, default=Path("data/derived_sample.csv"))
    return parser.parse_args()


def main() -> None:
    args = _parse_args()
    export_dataset(args.rows, args.seed, args.output)


if __name__ == "__main__":
    main()