Documentación de modelos

Ficha de modelos predictivos

Qué hace cada modelo, qué datos usa y cómo interpretar sus métricas.

Análisis de resultados

Modelos registrados

9 con datos en la tabla

Predicciones evaluables

107.017

115.784 registros historicos · 15.976 partidos base

Mejor acierto

Ensemble selectivo

76,77%

Última predicción

24/08/2026 08:51

Según `porcentajes_victoria.calculado_en`

Tabla general

9 modelos registrados. Pulsa un modelo para abrir su ficha técnica y sus datos actuales.

Todos Vista general 01 Elo básico 02 Reg. logística + Elo 03 Naive Bayes + Elo 04 Naive Bayes + stats 05 Ensemble jugadores 06 Ensemble calibrado 07 Ensemble selectivo 08 Comité especialistas 09 Modelo margen

Modelo	Tipo	Objetivo	Evaluables	Cobertura	Acierto	Última
Elo básico 01_elo_basico	Elo	Crear una probabilidad base usando fuerza historica de equipos.	15.927 49 neutras	100,00% útil 99,69%	62,55%	24/08/2026 08:26	Ver
Reg. logística + Elo 02_regresion_logistica_elo	Regresión logística	Mejorar Elo aprendiendo pesos estadisticos sobre variables dinamicas.	12.976 3.000 sin pred. histórica	81,22% útil 100,00%	63,98%	24/08/2026 08:26	Ver
Naive Bayes + Elo 03_naive_bayes_elo	Naive Bayes	Probar un clasificador probabilistico simple con señales Elo.	12.976 3.000 sin pred. histórica	81,22% útil 100,00%	62,48%	24/08/2026 08:26	Ver
Naive Bayes + stats 04_naive_bayes_stats	Naive Bayes + stats	Añadir estadisticas recientes de equipo al enfoque Naive Bayes.	12.976 3.000 sin pred. histórica	81,22% útil 100,00%	62,57%	24/08/2026 08:26	Ver
Ensemble jugadores 05_ensemble_jugadores	Ensemble	Combinar Elo, forma de equipo y perfiles rolling de jugadores.	12.975 1 neutras 3.000 sin pred. histórica	81,22% útil 99,99%	63,53%	24/08/2026 08:29	Ver
Ensemble calibrado 06_ensemble_calibrado	Ensemble calibrado	Mejorar la calibracion de probabilidades del ensemble.	11.976 4.000 sin pred. histórica	74,96% útil 100,00%	63,83%	24/08/2026 08:33	Ver
Ensemble selectivo 07_ensemble_selectivo	Ensemble selectivo	Subir el porcentaje de acierto seleccionando solo predicciones de alta confianza.	3.259 8.717 neutras 4.000 sin pred. histórica	74,96% útil 27,21%	76,77%	24/08/2026 08:39	Ver
Comité especialistas 08_comite_especialistas	Comité de especialistas	Mejorar el acierto usando un experto distinto según el tipo de partido.	11.976 4.000 sin pred. histórica	74,96% útil 100,00%	63,37%	24/08/2026 08:48	Ver
Modelo margen 09_modelo_margen_puntos	Modelo de margen	Predecir diferencia de puntos y convertirla en probabilidad de victoria.	11.976 4.000 sin pred. histórica	74,96% útil 100,00%	65,03%	24/08/2026 08:51	Ver

Modelo margen

09_modelo_margen_puntos

Modelo de margen

Qué hace

Cambia el objetivo: aprende margen local-visitante con regresores y calibra ese margen con regresión logística para obtener probabilidad de victoria local.

Resumen de datos que usa

Features completas del 05
Cuotas del 07
Target margen de puntos
Calibrador margen -> victoria

Validación resumida

Split temporal con accuracy, Brier, LogLoss, MAE y RMSE de margen.

Salida resumida

Probabilidad derivada del margen esperado.

Riesgo o lectura resumida

Puede mejorar información de confianza aunque no siempre suba el acierto bruto.

Tablas implicadas

Todas las del 05
cuota_partidos
porcentajes_victoria

Ficheros y responsabilidad

01_python/03_modelos/09_modelo_margen_puntos/modeling.py
Regresores de margen y calibración.
01_python/03_modelos/09_modelo_margen_puntos/01_train.py
Entrena el modelo de margen.
01_python/03_modelos/09_modelo_margen_puntos/03_backtest.py
Calcula histórico completo.
01_python/03_modelos/09_modelo_margen_puntos/02_apply_future.py
Aplica a futuros.

Features técnicas del JSON de entrenamiento

Numéricas (134)

diff_elo
abs_diff_elo
rest_local
rest_visitante
diff_rest
b2b_local
b2b_visitante
players_player_known_count_local
players_player_known_count_visitante
diff_players_player_known_count
players_player_roster_count_local
players_player_roster_count_visitante
diff_players_player_roster_count
players_player_top8_minutes_local
players_player_top8_minutes_visitante
diff_players_player_top8_minutes
players_player_top8_value_local
players_player_top8_value_visitante
diff_players_player_top8_value
players_player_top8_points_local
players_player_top8_points_visitante
diff_players_player_top8_points
players_player_top8_reb_local
players_player_top8_reb_visitante
diff_players_player_top8_reb
players_player_top8_ast_local
players_player_top8_ast_visitante
diff_players_player_top8_ast
players_player_top8_stocks_local
players_player_top8_stocks_visitante
diff_players_player_top8_stocks
players_player_top8_tov_local
players_player_top8_tov_visitante
diff_players_player_top8_tov
players_player_top8_ts_local
players_player_top8_ts_visitante
diff_players_player_top8_ts
players_player_top8_efg_local
players_player_top8_efg_visitante
diff_players_player_top8_efg
players_player_top8_usage_local
players_player_top8_usage_visitante
diff_players_player_top8_usage
players_player_top8_net_local
players_player_top8_net_visitante
diff_players_player_top8_net
players_player_top8_pm_local
players_player_top8_pm_visitante
diff_players_player_top8_pm
players_player_top8_pie_local
players_player_top8_pie_visitante
diff_players_player_top8_pie
players_player_depth_value_local
players_player_depth_value_visitante
diff_players_player_depth_value
players_player_depth_minutes_local
players_player_depth_minutes_visitante
diff_players_player_depth_minutes
team_win_rate_10_local
team_win_rate_10_visitante
diff_team_win_rate_10
team_margin_10_local
team_margin_10_visitante
diff_team_margin_10
team_points_for_10_local
team_points_for_10_visitante
diff_team_points_for_10
team_points_against_10_local
team_points_against_10_visitante
diff_team_points_against_10
team_team_fg_pct_10_local
team_team_fg_pct_10_visitante
diff_team_team_fg_pct_10
team_team_tp_pct_10_local
team_team_tp_pct_10_visitante
diff_team_team_tp_pct_10
team_team_ft_pct_10_local
team_team_ft_pct_10_visitante
diff_team_team_ft_pct_10
team_team_reb_10_local
team_team_reb_10_visitante
diff_team_team_reb_10
team_team_ast_10_local
team_team_ast_10_visitante
diff_team_team_ast_10
team_team_tov_10_local
team_team_tov_10_visitante
diff_team_team_tov_10
team_team_net_rating_10_local
team_team_net_rating_10_visitante
diff_team_team_net_rating_10
team_team_efg_10_local
team_team_efg_10_visitante
diff_team_team_efg_10
team_team_ts_10_local
team_team_ts_10_visitante
diff_team_team_ts_10
team_team_pace_10_local
team_team_pace_10_visitante
diff_team_team_pace_10
team_team_possessions_10_local
team_team_possessions_10_visitante
diff_team_team_possessions_10
team_team_pie_10_local
team_team_pie_10_visitante
diff_team_team_pie_10
odds_available
moneyline_local
moneyline_visitante
market_prob_local
market_prob_visitante
market_prob_diff
market_favorite_local
market_confidence
market_overround
spread_local
spread_visitante
total_puntos
elo_prob_local_proxy
diff_market_elo_prob
abs_diff_market_elo_prob
month
day_of_week
is_playoff_window
is_regular_season
home_rest_advantage
visitor_rest_advantage
both_b2b
favorite_alignment
market_elo_disagreement
strong_market_favorite
strong_elo_favorite
form_gap_abs
player_value_gap_abs

Categóricas (4)

tipo_partido
liga_id
equipo_local_id
equipo_visitante_id

Acierto

65,03%

Evaluables

11.976

Aciertos

7.788

Fallos

4.188

Cobertura histórica

74,96%

Cobertura útil

100,00%

Neutras 50/50

Sin pred. histórica

4.000

Omitidos OOF

4.000

Temporadas

Ligas

Periodo

03/12/2019 00:00
14/06/2026 00:30

Última predicción

24/08/2026 08:51

Métricas de entrenamiento

Modelo JSON

09_modelo_margen_puntos

Candidato seleccionado

extra_trees_margin

Total dataset

15.976

Train

10.864

Calibración

1.917

Test

3.195

Nombre candidato

extra_trees_margin

Brier test

0,2081

LogLoss test

0,6028

ROC AUC test

0,7236

Partidos con cuotas

5.500

Cobertura cuotas

34,43%

Cuotas en test

2.278

01_python/03_modelos/09_modelo_margen_puntos/metrics/modelo_margen_puntos_metrics.json

Walk-forward / OOF

Modelo OOF

09_modelo_margen_puntos

Candidato OOF

extra_trees_margin

Dataset total

15.976

Partidos backtest

11.976

Sin predicción OOF

4.000

Folds

Min train

3.000

Tamaño calibración

1.000

Tamaño fold

1.200

OOF accuracy @0.5

65,03%

OOF Brier

0,2171

OOF LogLoss

0,6234

OOF Winner ECE

—

OOF ROC AUC

0,6943

01_python/03_modelos/09_modelo_margen_puntos/metrics/modelo_margen_puntos_metrics_oof_metrics.json