""" Compara predicción del modelo GNN vs baseline vs realidad para Racing de Santander vs Sporting de Gijón y FC Andorra vs Racing de Santander. Re-ejecuta el modelo usando los mismos datos que tenía cuando se generaron los reportes head-to-head (dataset hasta marzo 2026). """ import sys import io import warnings import base64 from pathlib import Path from typing import Dict, List import numpy as np import pandas as pd import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt import matplotlib.patches as mpatches from matplotlib.gridspec import GridSpec warnings.filterwarnings("ignore") PROJECT_ROOT = Path(__file__).resolve().parents[1] sys.path.insert(0, str(PROJECT_ROOT / "scripts")) # Importar funciones del script de reportes from build_head_to_head_report_pro import ( _predict_single_team_future, _load_modeling_dataset, ModelPrediction, ) REPORTS_DIR = PROJECT_ROOT / "reports" REPORTS_DIR.mkdir(exist_ok=True) # ── Configuración ──────────────────────────────────────────────────────────── TEAM_IDS = { "Racing de Santander": "bzkwzatvwahmbzok1ymm5vqa1", "Sporting de Gijón": "9wfi6tgumbrnkp72z8zab89kr", "FC Andorra": "chhfxt372lm29p9b96fr2ho4q", } ZONE_ORDER = [ "Deep_Cross__Der_", "Half_Space__Der_", "Creativity_Zone", "Half_Space__Izq_", "Deep_Cross__Izq_", "Cross__Der_", "Cut_Back__Der_", "Scoring_Zone", "Cut_Back__Izq_", "Cross__Izq_", ] ZONE_LABELS = { "Deep_Cross__Der_": "ext der 3/4", "Half_Space__Der_": "int der 3/4", "Creativity_Zone": "centro 3/4", "Half_Space__Izq_": "int izq 3/4", "Deep_Cross__Izq_": "ext izq 3/4", "Cross__Der_": "ext der área", "Cut_Back__Der_": "int der área", "Scoring_Zone": "centro área", "Cut_Back__Izq_": "int izq área", "Cross__Izq_": "ext izq área", } ZONE_RECTS = { "Scoring_Zone": [(83, 100, 37, 63)], "Cut_Back__Izq_": [(83, 100, 63, 79)], "Cross__Izq_": [(83, 100, 79, 100)], "Cut_Back__Der_": [(83, 100, 21, 37)], "Cross__Der_": [(83, 100, 0, 21)], "Creativity_Zone": [(60, 83, 37, 63)], "Half_Space__Izq_": [(60, 83, 63, 79)], "Deep_Cross__Izq_": [(60, 83, 79, 100)], "Half_Space__Der_": [(60, 83, 21, 37)], "Deep_Cross__Der_": [(60, 83, 0, 21)], } COLORS = { "baseline": "#5B8DB8", "model": "#F0B429", "actual": "#E05A2B", } DARK_BG = "#1A1A2E" PANEL_BG = "#16213E" TEXT_COL = "#E8E8E8" GRID_COL = "#2E4057" # ── Datos reales desde eventos ─────────────────────────────────────────────── def assign_attack_zone(df: pd.DataFrame) -> pd.Series: zone_col = pd.Series(index=df.index, dtype="object") x = df["x"].astype(float) y = df["y"].astype(float) for zone_name, rects in ZONE_RECTS.items(): mask = pd.Series(False, index=df.index) for (x0, x1, y0, y1) in rects: mask |= x.ge(x0) & x.lt(x1) & y.ge(y0) & y.lt(y1) zone_col.mask(mask, zone_name, inplace=True) return zone_col def compute_actual(events_df: pd.DataFrame, match_id: str, team_id: str) -> Dict[str, Dict[str, float]]: df = events_df[ (events_df["matchId"] == match_id) & (events_df["teamId"] == team_id) ].copy() if df.empty: return {"attack": {z: 0.0 for z in ZONE_ORDER}, "pv": {z: 0.0 for z in ZONE_ORDER}} if "outcome_value" in df.columns: df_z = df[(df["outcome_value"] == 1) & df["x"].notna() & df["y"].notna()].copy() else: df_z = df[df["x"].notna() & df["y"].notna()].copy() df_z["attack_zone"] = assign_attack_zone(df_z) df_z = df_z[df_z["attack_zone"].notna()] # Attack share counts = df_z.groupby("attack_zone").size().reindex(ZONE_ORDER, fill_value=0).astype(float) total = counts.sum() attack = (counts / total).to_dict() if total > 0 else {z: 0.0 for z in ZONE_ORDER} # PV share df_z["pvAdded"] = pd.to_numeric(df_z["pvAdded"], errors="coerce").fillna(0) pv_by_zone = df_z[df_z["pvAdded"] > 0].groupby("attack_zone")["pvAdded"].sum().reindex(ZONE_ORDER, fill_value=0) pv_total = pv_by_zone.sum() pv = (pv_by_zone / pv_total).to_dict() if pv_total > 0 else {z: 0.0 for z in ZONE_ORDER} return {"attack": attack, "pv": pv} # ── Gráficos ───────────────────────────────────────────────────────────────── def _setup_ax(ax): ax.set_facecolor(PANEL_BG) for spine in ax.spines.values(): spine.set_color(GRID_COL) ax.tick_params(colors=TEXT_COL, labelsize=7) ax.yaxis.grid(True, color=GRID_COL, linewidth=0.5, zorder=0) ax.set_axisbelow(True) def make_triple_bar(ax, zones, baseline_vals, model_vals, actual_vals, title, ylabel): _setup_ax(ax) x = np.arange(len(zones)) w = 0.26 bl = [baseline_vals.get(z, 0) * 100 for z in zones] mo = [model_vals.get(z, 0) * 100 for z in zones] ac = [actual_vals.get(z, 0) * 100 for z in zones] b1 = ax.bar(x - w, bl, width=w, color=COLORS["baseline"], alpha=0.85, label="Baseline histórico", zorder=3) b2 = ax.bar(x, mo, width=w, color=COLORS["model"], alpha=0.90, label="Modelo GNN", zorder=3) b3 = ax.bar(x + w, ac, width=w, color=COLORS["actual"], alpha=0.90, label="Real (partido)", zorder=3) ax.set_xticks(x) ax.set_xticklabels([ZONE_LABELS[z] for z in zones], fontsize=7.5, color=TEXT_COL, rotation=30, ha="right") ax.set_ylabel(ylabel, color=TEXT_COL, fontsize=8) ax.set_title(title, color=TEXT_COL, fontsize=9, fontweight="bold", pad=6) ax.set_ylim(0, max(max(bl + mo + ac) * 1.35, 5)) leg = ax.legend(handles=[b1, b2, b3], fontsize=7, facecolor=PANEL_BG, edgecolor=GRID_COL, labelcolor=TEXT_COL, framealpha=0.85, loc="upper right") def make_error_panel(ax, zones, baseline_vals, model_vals, actual_vals, title): """ Error absoluto por zona: |modelo - real| vs |baseline - real| Barra verde = modelo ganó (se acercó más), roja = baseline ganó. """ _setup_ax(ax) err_model = [abs(model_vals.get(z, 0) - actual_vals.get(z, 0)) * 100 for z in zones] err_baseline = [abs(baseline_vals.get(z, 0) - actual_vals.get(z, 0)) * 100 for z in zones] diff = [eb - em for em, eb in zip(err_model, err_baseline)] # + = modelo mejor colors = ["#2ECC71" if d >= 0 else "#E74C3C" for d in diff] x = np.arange(len(zones)) bars = ax.bar(x, diff, color=colors, alpha=0.88, zorder=3) for bar, d in zip(bars, diff): if abs(d) > 0.2: va = "bottom" if d >= 0 else "top" ax.text(bar.get_x() + bar.get_width()/2, bar.get_height(), f"{'+' if d > 0 else ''}{d:.1f}pp", ha="center", va=va, fontsize=6, color=TEXT_COL, fontweight="bold") ax.axhline(0, color=TEXT_COL, linewidth=0.8, zorder=2) ax.set_xticks(x) ax.set_xticklabels([ZONE_LABELS[z] for z in zones], fontsize=7.5, color=TEXT_COL, rotation=30, ha="right") ax.set_ylabel("Ventaja modelo (pp)\nError baseline − Error modelo", color=TEXT_COL, fontsize=7.5) ax.set_title(title, color=TEXT_COL, fontsize=9, fontweight="bold", pad=6) lim = max(abs(d) for d in diff) * 1.4 if any(diff) else 5 ax.set_ylim(-lim, lim) # Mini leyenda p_green = mpatches.Patch(color="#2ECC71", label="Modelo más preciso") p_red = mpatches.Patch(color="#E74C3C", label="Baseline más preciso") ax.legend(handles=[p_green, p_red], fontsize=7, facecolor=PANEL_BG, edgecolor=GRID_COL, labelcolor=TEXT_COL, framealpha=0.85, loc="upper right") def compute_summary(zones, baseline_vals, model_vals, actual_vals, metric): """Devuelve MAE del modelo y del baseline, y en cuántas zonas ganó cada uno.""" err_model = [abs(model_vals.get(z, 0) - actual_vals.get(z, 0)) for z in zones] err_baseline = [abs(baseline_vals.get(z, 0) - actual_vals.get(z, 0)) for z in zones] mae_model = np.mean(err_model) * 100 mae_baseline = np.mean(err_baseline) * 100 zones_model = sum(1 for em, eb in zip(err_model, err_baseline) if em < eb) zones_base = len(zones) - zones_model winner = "Modelo" if mae_model < mae_baseline else "Baseline" return { "metric": metric, "mae_model": mae_model, "mae_baseline": mae_baseline, "zones_model": zones_model, "zones_base": zones_base, "winner": winner, } def fig_to_b64(fig) -> str: buf = io.BytesIO() fig.savefig(buf, format="png", dpi=150, bbox_inches="tight", facecolor=DARK_BG) plt.close(fig) buf.seek(0) return base64.b64encode(buf.read()).decode() def build_match_figure(match_title, teams_data) -> str: """ Genera figura completa para un partido: 2 equipos x 3 filas (ataque, pv, error). """ team_names = list(teams_data.keys()) n = len(team_names) fig = plt.figure(figsize=(16, 15), facecolor=DARK_BG) fig.suptitle(match_title, color=TEXT_COL, fontsize=13, fontweight="bold", y=0.99) gs = GridSpec(3, n, figure=fig, hspace=0.72, wspace=0.32, top=0.94, bottom=0.07) summaries = [] for col, team_name in enumerate(team_names): d = teams_data[team_name] bl = d["baseline"] mo = d["model"] ac = d["actual"] make_triple_bar( ax=fig.add_subplot(gs[0, col]), zones=ZONE_ORDER, baseline_vals=bl["attack"], model_vals=mo["attack"], actual_vals=ac["attack"], title=f"{team_name} — Share de Ataque", ylabel="% acciones en zona", ) make_triple_bar( ax=fig.add_subplot(gs[1, col]), zones=ZONE_ORDER, baseline_vals=bl["pv"], model_vals=mo["pv"], actual_vals=ac["pv"], title=f"{team_name} — Share de Peligro (pvAdded)", ylabel="% pvAdded en zona", ) make_error_panel( ax=fig.add_subplot(gs[2, col]), zones=ZONE_ORDER, baseline_vals=bl["attack"], model_vals=mo["attack"], actual_vals=ac["attack"], title=f"{team_name} — Ventaja del modelo vs baseline (Ataque)", ) summaries.append(compute_summary(ZONE_ORDER, bl["attack"], mo["attack"], ac["attack"], f"{team_name} Ataque")) summaries.append(compute_summary(ZONE_ORDER, bl["pv"], mo["pv"], ac["pv"], f"{team_name} Peligro")) # Leyenda global patches = [ mpatches.Patch(color=COLORS["baseline"], label="Baseline (promedio histórico)"), mpatches.Patch(color=COLORS["model"], label="Modelo GNN (predicción pre-partido)"), mpatches.Patch(color=COLORS["actual"], label="Real (lo que ocurrió)"), mpatches.Patch(color="#2ECC71", label="Zona donde el modelo ganó al baseline"), mpatches.Patch(color="#E74C3C", label="Zona donde el baseline fue mejor"), ] fig.legend(handles=patches, loc="lower center", ncol=3, fontsize=8, facecolor=PANEL_BG, edgecolor=GRID_COL, labelcolor=TEXT_COL, framealpha=0.9, bbox_to_anchor=(0.5, 0.01)) return fig_to_b64(fig), summaries def summary_table_html(summaries: list) -> str: rows = "" for s in summaries: winner_color = "#F0B429" if s["winner"] == "Modelo" else "#5B8DB8" rows += f""" {s['metric']} {s['mae_baseline']:.2f} pp {s['mae_model']:.2f} pp {s['winner']} {s['zones_model']}/10 zonas """ return f""" {rows}
Métrica MAE Baseline MAE Modelo GNN Ganador Zonas modelo mejor
""" def build_html(matches: list, output_path: Path): cards = "" all_summaries = [] for m in matches: img_b64, summaries = build_match_figure(m["title"], m["teams"]) all_summaries.extend(summaries) cards += f"""

{m['title']}

""" table_html = summary_table_html(all_summaries) html = f""" Modelo GNN vs Baseline vs Realidad — Racing
Post-partido · Análisis de predicción

Modelo GNN vs Baseline vs Realidad

Comparación de las predicciones del modelo GNN y el baseline histórico contra lo que realmente ocurrió en los partidos de Racing de Santander. El panel inferior de cada equipo muestra la ventaja del modelo sobre el baseline en cada zona: verde = modelo más cercano a la realidad, rojo = baseline fue mejor.

Baseline: Promedio histórico de la temporada para cada equipo.
Modelo GNN: Predicción generada antes del partido usando el dataset de modelado (hasta 08/03/2026).
Real: Distribución observada en el partido, calculada desde los eventos crudos.
MAE: Error Absoluto Medio en puntos porcentuales (pp). Menor = mejor.
{cards}

Resumen global — ¿Quién ganó?

{table_html}
""" output_path.write_text(html, encoding="utf-8") print(f"Reporte generado: {output_path}") # ── Main ───────────────────────────────────────────────────────────────────── def main(): print("Cargando dataset de modelado...") df_model = _load_modeling_dataset() print("Cargando eventos crudos...") events_df = pd.read_csv( "/Users/pagrois/Documents/Racing/preprocessed_SSD_25-26.csv", usecols=["matchId", "teamId", "x", "y", "outcome_value", "pvAdded"], dtype={"matchId": str, "teamId": str}, low_memory=True, ) LEAGUE = "Spanish Segunda Division" SEASON = "25-26" MATCHES = [ { "title": "Racing de Santander vs Sporting de Gijón — 1 de abril 2026", "match_id": "7hegc9covicy699bxsi81xkb8", "home": ("Racing de Santander", "bzkwzatvwahmbzok1ymm5vqa1"), "away": ("Sporting de Gijón", "9wfi6tgumbrnkp72z8zab89kr"), }, { "title": "FC Andorra vs Racing de Santander — 5 de abril 2026", "match_id": "7n8819yv16f6hm7xt95007bis", "home": ("FC Andorra", "chhfxt372lm29p9b96fr2ho4q"), "away": ("Racing de Santander", "bzkwzatvwahmbzok1ymm5vqa1"), }, ] matches_out = [] for mc in MATCHES: print(f"\nProcesando: {mc['title']}") teams_out = {} for role, is_home in [("home", True), ("away", False)]: team_name, team_id = mc[role] opp_name, opp_id = mc["away" if role == "home" else "home"] print(f" → Predicción modelo: {team_name}...") pred: ModelPrediction = _predict_single_team_future( df_model=df_model, team_id=team_id, opponent_team_id=opp_id, team_name=team_name, opponent_name=opp_name, league=LEAGUE, season=SEASON, is_home=is_home, ) print(f" → Datos reales: {team_name}...") actual = compute_actual(events_df, mc["match_id"], team_id) teams_out[team_name] = { "baseline": { "attack": pred.attack_baseline, "pv": pred.pv_baseline, }, "model": { "attack": pred.attack, "pv": pred.pv, }, "actual": actual, } matches_out.append({"title": mc["title"], "teams": teams_out}) output = REPORTS_DIR / "modelo_vs_realidad_Racing.html" print("\nGenerando HTML...") build_html(matches_out, output) return output if __name__ == "__main__": out = main() print(f"\nListo: {out}")