temporal_estrat_train_test.py

def temporal_train_test_split(df, predictors, target, test_size=0.2):
    """
    División temporal que respeta:
    1. Orden cronológico (no leakage)
    2. Integridad de series por país (no partir series arbitrariamente)
    3. Balance de países entre train/test
    """
    # Ordenar por año
    df_sorted = df.sort_values('Year').reset_index(drop=True)
    
    # Determinar punto de corte temporal
    cutoff_year = df_sorted['Year'].quantile(1 - test_size)
    
    train = df_sorted[df_sorted['Year'] <= cutoff_year]
    test = df_sorted[df_sorted['Year'] > cutoff_year]
    
    # Verificar que no haya leakage de países
    countries_train = set(train['Country'].unique())
    countries_test = set(test['Country'].unique())
    countries_overlap = countries_train & countries_test
    
    print(f"División temporal:")
    print(f"  - Año de corte: {cutoff_year:.0f}")
    print(f"  - Train: {len(train)} obs ({train['Year'].min()}-{train['Year'].max()})")
    print(f"  - Test:  {len(test)} obs ({test['Year'].min()}-{test['Year'].max()})")
    print(f"  - Países en ambos sets: {len(countries_overlap)} (evitar overfitting)")
    
    return (
        train[predictors].values,
        train[target].values,
        test[predictors].values,
        test[target].values,
        train, test
    )