oracle-ai-developer-hub

ML Feature Pipeline Notes

This workshop’s predictions come from the same enhanced Reto Enseña 4.0 pipeline used in enhanced_features.py, notebooks/world_cup_enhanced.ipynb, and the runtime FeatureRuntime tools. The final model artifact (models/best_model.pkl) declares 92 features, and scripts/verify.py checks that the artifact still exposes all 92 before the workshop is considered ready.

Data flow at a glance

1. Canonical input data: results.csv, goalscorers.csv, and shootouts.csv from the Kaggle international football results dataset.
2. Oracle load: scripts/setup_db.py loads the CSVs into MATCH_RESULTS, GOALSCORERS, SHOOTOUTS, and analytical views such as VW_TEAM_STATISTICS.
3. Chronological replay: every tracker reads only pre-match state before a row is emitted, then updates after the result. This is the main data-leakage guard.
4. 92-feature row: feature families are concatenated into one numeric row per match.
5. Model training/inference: the artifact is a three-class XGBoost-style classifier for Win, Draw, and Loss probabilities. predict_match builds the same feature row live for hypothetical/current matchups.
6. Post-inference retrieval: bulk predictions are loaded into PREDICCIONES_FINAL, then scripts/load_langchain_vectors.py turns predictions and football facts into LangChain OracleVS documents in SOCCER_LANGCHAIN_DOCS using langchain-oracledb.

Elo scoring system

The Elo tracker starts every team at 1500, replays matches chronologically, and maintains both a global rating and tournament-category-specific ratings.

Expected score

E_home = 1 / (1 + 10^(-(R_home - R_away + HA) / 400))

• R_home, R_away: current pre-match ratings.
• HA: home-advantage bonus, 100 Elo points on non-neutral venues and 0 on neutral venues.
• 400: standard Elo scaling constant.

Rating update

R_new = R_old + K * G * (S_actual - S_expected)

• S_actual: 1.0 for a win, 0.5 for a draw, 0.0 for a loss.
• K: match-importance factor.
• G: goal-difference multiplier.

K-factors

Tournament category	K-factor	Notes
FIFA World Cup	60	Highest-stakes international matches.
Continental championship	50	Euros, Copa América, AFCON, Asian Cup, Gold Cup, Confederations Cup, etc.
Qualifier / Nations League / default competitive	40	Competitive but below finals-stage intensity.
Friendly	20	Lowest weight; lineups and intensity are less reliable.

Goal-difference multiplier

Goal difference	Multiplier
0–1	1.0
2	1.5
3	1.75
4+	1.75 + (GD - 3) / 8

Large margins matter, but with diminishing returns.

Elo-derived features — 8

Feature	Meaning
home_elo, away_elo	Global pre-match ratings.
elo_diff, elo_total	Rating gap and combined match quality proxy.
home_expected	Elo-implied home result expectation.
home_tournament_elo, away_tournament_elo	Category-specific Elo ratings.
tournament_elo_diff	Tournament-category rating gap.

Feature families in the final 92-feature model

Family	Count	Source	What it captures
Original baseline: Elo, form/goals, H2H, context	40	results.csv	Long-term quality, recent trajectory, pair history, venue/tournament/rest context.
Goalscorer intelligence	12	goalscorers.csv	How teams score: scoring depth, star dependency, penalties, late goals, first-half share.
Momentum / psychology	16	results.csv + goalscorers.csv	Streaks, unbeaten runs, clean sheets, comebacks, draw tendency, blowout/shutout behavior.
Poisson expected goals	8	results.csv	Goal-rate lambdas, Poisson win/draw probabilities, scoring variance, over/underperformance.
Venue / geography	5	match metadata + lookup tables	Altitude, confederation effects, intercontinental matchup flag.
Tournament context	11	results.csv	World Cup form, competitive form, big-game factor, World Cup experience.
Total	92

Exact feature names by family

Original baseline — 40

• Elo: home_elo, away_elo, elo_diff, elo_total, home_tournament_elo, away_tournament_elo, tournament_elo_diff, home_expected.
• Form: home_form_5, home_form_10, home_form_20, away_form_5, away_form_10, away_form_20, home_weighted_form_10, away_weighted_form_10, form_diff_5, form_diff_10, weighted_form_diff.
• Goal statistics: home_goals_scored_avg_10, home_goals_conceded_avg_10, away_goals_scored_avg_10, away_goals_conceded_avg_10, home_goal_diff_avg_10, away_goal_diff_avg_10, goal_diff_differential, attack_vs_defense.
• Head-to-head: h2h_win_rate, h2h_matches, h2h_goal_diff.
• Match context: is_neutral, is_home, is_world_cup, is_continental, is_friendly, home_days_rest, away_days_rest, rest_diff, home_experience, away_experience.

Goalscorer intelligence — 12

home_scoring_depth, away_scoring_depth, scoring_depth_diff, home_star_dependency, away_star_dependency, home_penalty_ratio, away_penalty_ratio, home_late_goal_ratio, away_late_goal_ratio, late_goal_diff, home_first_half_ratio, away_first_half_ratio.

Momentum / psychology — 16

home_streak, away_streak, streak_diff, home_unbeaten, away_unbeaten, home_clean_sheet_pct, away_clean_sheet_pct, home_comeback_rate, away_comeback_rate, home_draw_tendency, away_draw_tendency, draw_tendency_sum, home_blowout_win_pct, away_blowout_loss_pct, home_shutout_loss_pct, away_shutout_loss_pct.

Poisson expected goals — 8

home_lambda, away_lambda, home_poisson_win, home_poisson_draw, home_scoring_variance, away_scoring_variance, home_overperformance, away_overperformance.

Venue / geography — 5

altitude, is_high_altitude, same_confederation, confed_strength_diff, is_intercontinental.

Tournament context — 11

home_wc_form, away_wc_form, wc_form_diff, home_competitive_form, away_competitive_form, home_big_game_factor, away_big_game_factor, big_game_diff, home_wc_experience, away_wc_experience, wc_experience_diff.

Why the agent tools mirror the feature families

The final chat agent exposes the same pipeline as tools so a user can ask for both the prediction and the evidence behind it:

Runtime tool	Feature-family source
get_elo	FootballElo — global and tournament Elo.
get_team_form	TeamTracker — rolling form, weighted form, goals.
get_h2h	H2HTracker — pair-specific history.
get_momentum	MomentumTracker — psychological/momentum features.
get_poisson_xg	PoissonTracker — statistical goal model.
get_tournament_context	TournamentTracker — big-game and World Cup context.
predict_match	Builds all 92 features and runs live inference.
hybrid_retrieve	Retrieves post-inference prediction/fact documents from the LangChain OracleVS store.

The important workshop claim is therefore falsifiable: a final answer can show the live 92-feature prediction and retrieve the corresponding cached prediction/fact evidence from Oracle through langchain-oracledb.