"""Page 6 — Dev Preview: New ML Models for Auction Optimization. Showcases all 12 new ML modules with real 26/27 Serie A data. """ import sys from pathlib import Path _p = Path(__file__).resolve().parent.parent.parent _str_p_ = str(_p) if _str_p_ not in sys.path: sys.path.insert(0, _str_p_) import numpy as np import pandas as pd import streamlit as st import plotly.graph_objects as go from plotly.subplots import make_subplots from dashboard.viz.components import inject_css, section, insight, role_chip, kpi_card from dashboard.viz.template import ( PITCH_GREEN, GOLD, RED, SKY, VIOLET, BG, CARD_BG, BORDER, TEXT, TEXT_SECONDARY, WHITE, ROLE_COLORS, ROLE_ICONS, FANTABETO_TEMPLATE, HEATMAP_COLORS, GRIDLINE, ) DATA_ROOT = Path(__file__).resolve().parent.parent.parent # ─── Real data loading ────────────────────────────────────────────── @st.cache_data(ttl=3600) def _load_real_data(): projections = pd.read_excel(DATA_ROOT / "data" / "player_projections_26_27.xlsx") stats = pd.read_excel(DATA_ROOT / "mid_outputs" / "players_stats.xlsx") votes = pd.read_excel(DATA_ROOT / "mid_outputs" / "players_votes.xlsx") roster = pd.read_excel(DATA_ROOT / "fantacalcio" / "Quotazioni_Fantacalcio_26_27.xlsx") auction_plan = pd.read_excel(DATA_ROOT / "data" / "auction_plan_al_cihred.xlsx") return projections, stats, votes, roster, auction_plan def _build_rich_player_pool(projections, stats): p = projections.copy() s = stats.copy() p["name_lower"] = p["player"].str.lower().str.strip() s["name_lower"] = s["name"].str.lower().str.strip() # Merge stats onto projections by name stat_features = [ "minutes", "goals_p90", "assists_p90", "xg_per90", "npxg_per90", "xa_per90", "shots_on_target_pct", "passes_pct", "progressive_passes", "progressive_carries", "tackles", "interceptions", "clearances", "aerials_won_pct", "fouls", "fouled", "cards_yellow", "cards_red", "sca_per90", "gca_per90", "touches_att_3rd", "touches_att_pen_area", "passes_into_final_third", "crosses_into_penalty_area", "gk_save_pct", "gk_clean_sheets_pct", "gk_psxg_net_per90", ] available = [c for c in stat_features if c in s.columns] merge_cols = ["name_lower"] + available merged = p.merge(s[merge_cols], on="name_lower", how="left") for c in available + ["fv_std", "qi", "goals", "assists", "cards_yellow", "cards_red", "fouls", "fouled", "xg_per90", "xa_per90", "sca_per90", "gca_per90", "minutes", "starter_pct", "games"]: if c in merged.columns: merged[c] = merged[c].fillna(0) merged["rest_days"] = np.random.RandomState(42).uniform(2, 10, len(merged)) merged["fatigue_rolling_3"] = (merged["minutes"] * 0.33).clip(0, 90) merged["goals_season"] = merged["goals"] merged["assists_season"] = merged["assists"] minutes = merged["minutes"].clip(lower=1) merged["yellow_per_game"] = (merged["cards_yellow"] / (minutes / 90)).clip(0, 1) merged["red_per_game"] = (merged["cards_red"] / (minutes / 90)).clip(0, 0.5) merged["fouls_p90"] = merged["fouls"] / (minutes / 90) merged["fouled_p90"] = merged["fouled"] / (minutes / 90) merged["projected_points"] = merged["fv_proj"].fillna(6.0) merged["fv_std"] = merged["fv_std"].fillna(0.5) merged["market_value"] = merged["qi"] merged["games_played"] = merged["games"] merged["name"] = merged["player"] return merged def _build_vote_features(votes): if "fantavote" not in votes.columns: return votes vote_avg = votes.groupby("player").agg( vote_avg=("fantavote", "mean"), vote_std=("fantavote", "std"), vote_count=("fantavote", "count"), ).reset_index() return vote_avg def _generate_interaction_data(player_pool): """Generate synthetic interactions scaled by real stats.""" rng = np.random.RandomState(42) rows = [] players = player_pool["name"].tolist() teams = player_pool["team"].tolist() player_team = dict(zip(players, teams)) for _ in range(800): a = rng.choice(players) b = rng.choice(players) if a == b: continue same_team = 1.5 if player_team.get(a) == player_team.get(b) else 0.2 rows.append({ "player": a, "teammate": b, "passes_to": int(rng.exponential(3 * same_team)), "assists_to": int(rng.exponential(0.3 * same_team)), "crosses_to": int(rng.exponential(1 * same_team)), "matchday": rng.randint(1, 39), }) return pd.DataFrame(rows) def _generate_auction_logs(player_pool): rng = np.random.RandomState(42) rows = [] for _ in range(1000): player = player_pool.iloc[rng.randint(0, len(player_pool))] points = player["projected_points"] qi = player["market_value"] rows.append({ "player_name": player["name"], "player_role": player["role"], "player_projected_points": points, "opponent_budget_remaining": rng.uniform(100, 500), "opponent_slots_remaining": rng.randint(1, 8), "role_needed_count": rng.randint(1, 5), "round_number": rng.randint(1, 15), "winning_bid": max(1, int(qi * rng.uniform(0.5, 2.2))), }) return pd.DataFrame(rows) def _generate_team_rosters(player_pool, n_teams=8): rng = np.random.RandomState(42) rosters = [] for t in range(n_teams): idx = rng.choice(len(player_pool), 25, replace=False) roster = player_pool.iloc[idx].copy() roster["team"] = f"Team_{t}" rosters.append(roster) return rosters # ─── Model initialization ─────────────────────────────────────────── @st.cache_resource def _init_models(player_pool, interaction_data, auction_logs): results = {} feature_cols = ["projected_points", "fv_std", "games_played", "starter_pct", "goals_season", "assists_season", "yellow_per_game", "red_per_game", "xg_per90", "xa_per90", "sca_per90", "gca_per90", "fouls_p90", "minutes", "rest_days", "fatigue_rolling_3"] feature_cols = [c for c in feature_cols if c in player_pool.columns] X_full = player_pool[feature_cols].fillna(0) y_full = player_pool["projected_points"].values # 1 ─ Quantile Ensemble try: from src.models.quantile_model import QuantileEnsemble qe = QuantileEnsemble(quantiles=(0.10, 0.50, 0.90), n_estimators=100) qe.fit(X_full, pd.Series(y_full)) top60 = player_pool.nlargest(60, "projected_points") X_top = top60[feature_cols].fillna(0) preds = qe.predict(X_top) risk = qe.predict_downside_risk(X_top, threshold=5.5) results["quantile"] = {"preds": {k: v.tolist() for k, v in preds.items()}, "risk": risk.tolist(), "players": top60["name"].tolist()} except Exception as e: results["quantile"] = {"error": str(e)} # 2 ─ Survival Model try: from src.models.survival_model import MinutesSurvivalModel surv_features = ["minutes", "games_played", "rest_days", "fatigue_rolling_3", "starter_pct"] surv_features = [c for c in surv_features if c in player_pool.columns] surv_df = player_pool[surv_features].fillna(0) durations = np.clip(player_pool["minutes"].fillna(60).values, 1, 90) events = (player_pool["starter_pct"].fillna(0.5).values > 0.5).astype(int) ms = MinutesSurvivalModel(force_scipy=True) ms.fit(surv_df, durations, events) sample = surv_df.head(40) expected, lower, upper = ms.predict_distribution(sample) starter_probs = ms.predict_starter_probability(sample) results["survival"] = {"expected": expected.tolist(), "lower": lower.tolist(), "upper": upper.tolist(), "starter_probs": starter_probs.tolist()} except Exception as e: results["survival"] = {"error": str(e)} # 3 ─ Bayesian Pooling try: from src.models.bayesian_pooling import BayesianPlayerModel Xb = player_pool[["role", "projected_points", "games_played", "starter_pct"]].head(200).copy() yb = player_pool["projected_points"].head(200) bp = BayesianPlayerModel() bp.fit(Xb, yb) top30 = player_pool.nlargest(30, "projected_points") Xb30 = top30[["role", "projected_points", "games_played", "starter_pct"]].copy() mean, std = bp.predict_with_uncertainty(Xb30) reliability = bp.get_player_reliability(Xb30) results["bayesian"] = {"mean": mean.tolist(), "std": std.tolist(), "reliability": reliability.tolist(), "players": top30["name"].tolist()} except Exception as e: results["bayesian"] = {"error": str(e)} # 4 ─ Conformal Predictor try: from src.models.conformal_predictor import ConformalPredictor from sklearn.linear_model import Ridge Xcp = X_full.head(300).fillna(0) ycp = y_full[:300] base = Ridge(alpha=1.0) base.fit(Xcp.iloc[:200], ycp[:200]) cp = ConformalPredictor(base, alpha=0.10) cp.calibrate(Xcp.iloc[200:250], pd.Series(ycp[200:250])) yp, yl, yu = cp.predict_with_band(Xcp.iloc[250:270]) coverage = cp.coverage(Xcp.iloc[250:270], pd.Series(ycp[250:270])) results["conformal"] = {"coverage": float(coverage), "n_test": 20, "predictions": yp.tolist(), "lowers": yl.tolist(), "uppers": yu.tolist()} except Exception as e: results["conformal"] = {"error": str(e)} # 5 ─ Bandit try: from src.optimization.bandit_auction import BanditAuctionSolver from src.optimization.auction_solver import AuctionConfig, PlayerValuation config = AuctionConfig(total_budget=500, n_gk=3, n_def=8, n_mid=8, n_fwd=6) bandit = BanditAuctionSolver(config=config) trial_bids = [] for i in range(30): player = player_pool.iloc[i] pv = PlayerValuation( name=player["name"], team=player["team"], role=player["role"], projected_points=player["projected_points"], market_value=player["market_value"], ceiling_price=player["projected_points"] * 5, ) state = { "budget_remaining": max(50, 500 - i * 15), "total_budget": 500, "slots_remaining": {"P": max(0, 1 - i//30), "D": max(0, 3 - i//10), "C": max(0, 4 - i//7), "A": max(0, 2 - i//15)}, "role_quotas": {"P": 3, "D": 8, "C": 8, "A": 6}, "slot_quotas": {"P": 3, "D": 8, "C": 8, "A": 6}, "round_number": i + 1, "total_rounds": 30, "opponent_budgets": [400, 350, 420], "players_remaining_in_role": {"P": 15, "D": 50, "C": 50, "A": 30}, "player_pool": [pv], } arm, bid = bandit.select_bid(pv, state) bandit.update(arm, 0.6, pv.role) trial_bids.append({"player": pv.name, "role": pv.role, "bid": bid, "arm": arm}) results["bandit"] = {"trial_bids": trial_bids, "arm_stats": bandit.get_arm_stats()} except Exception as e: results["bandit"] = {"error": str(e)} # 6 ─ Opponent Bidding try: from src.optimization.opponent_bidding_model import OpponentBidModel obm = OpponentBidModel() sample_players = player_pool.head(50).rename(columns={ "name": "player_name", "role": "player_role", "projected_points": "player_projected_points", }) opp_state = { "budget_remaining": 400, "total_budget": 500, "initial_budget": 500, "slots_remaining": {"P": 2, "D": 6, "C": 6, "A": 4}, "slots_total": {"P": 3, "D": 8, "C": 8, "A": 6}, "slots_filled": {"P": 1, "D": 2, "C": 2, "A": 2}, "role_quotas": {"P": 3, "D": 8, "C": 8, "A": 6}, "aggression_factor": 1.0, } bids = obm.predict_opponent_bids(sample_players, opp_state) results["opponent_bidding"] = { "sample_bids": bids.head(30).tolist(), "players": sample_players["player_name"].head(30).tolist(), } except Exception as e: results["opponent_bidding"] = {"error": str(e)} # 7 ─ Budget Optimizer try: from src.optimization.budget_optimizer import BudgetOptimizer bo = BudgetOptimizer(total_budget=500) allocation = bo.optimize(player_pool, n_calls=15) curves = bo.get_role_value_curves(player_pool) results["budget_opt"] = {"allocation": allocation, "curves": curves} except Exception as e: results["budget_opt"] = {"error": str(e)} # 8 ─ GAT Chemistry try: from src.models.gat_model import PlayerChemistryGAT gat = PlayerChemistryGAT() gat.build_graph(interaction_data) top_players = player_pool.nlargest(6, "projected_points")["name"].tolist() bonus_matrix = [] for p in top_players: row = [] for q in top_players: row.append(gat.compute_interaction_bonus(p, q)) bonus_matrix.append(row) chem_features = gat.extract_interaction_features(top_players[0], top_players) results["chemistry"] = {"features": chem_features, "bonus_matrix": bonus_matrix, "labels": [n.split()[-1] for n in top_players]} except Exception as e: results["chemistry"] = {"error": str(e)} # 9 ─ Hawkes Form try: from src.models.hawkes_form import PlayerFormModel import importlib as _il _self = _il.import_module('dashboard.pages.06_dev_preview') vote_avg = _self._build_vote_features(pd.read_excel(DATA_ROOT / "mid_outputs" / "players_votes.xlsx")) form_players = player_pool.head(30).copy() form_players["name_lower"] = form_players["name"].str.lower() vote_avg["name_lower"] = vote_avg["player"].str.lower() form_merged = form_players.merge(vote_avg[["name_lower", "vote_avg", "vote_std", "vote_count"]], on="name_lower", how="left") dates = pd.date_range("2025-08-20", periods=len(form_merged), freq="7D") form_data = pd.DataFrame({ "player": form_merged["name"], "match_date": dates, "minutes": form_merged["minutes"].fillna(60), }) form_y = form_merged["projected_points"] hf = PlayerFormModel() hf.fit(form_data, form_y) statuses = hf.get_form_status(form_data) results["hawkes"] = {"players": form_merged["name"].tolist()[:15], "statuses": {form_merged["name"].iloc[i]: statuses[i] for i in range(min(15, len(statuses)))}} except Exception as e: results["hawkes"] = {"error": str(e)} # 10 ─ RL Agent try: from src.optimization.rl_auction_agent import AuctionEnv, RLAuctionPolicy, RLAuctionTrainer from src.optimization.auction_solver import AuctionConfig config_small = AuctionConfig(total_budget=500, n_gk=1, n_def=3, n_mid=3, n_fwd=2) env = AuctionEnv(player_pool.head(40), n_opponents=2, config=config_small) agent = RLAuctionPolicy(state_dim=14, action_dim=11, hidden_dim=32) trainer = RLAuctionTrainer(player_pool.head(40), n_opponents=2, config=config_small) agent = trainer.train(n_episodes=30, verbose=False) rl_state = env.reset() action = agent.act(rl_state, epsilon=0.0) results["rl"] = {"observation_dim": len(rl_state), "action_taken": int(action), "action_dim": 11, "buffer_size": len(agent.replay_buffer)} except Exception as e: results["rl"] = {"error": str(e)} # 11 ─ Set Transformer try: from src.models.set_transformer import SetTransformer rosters = _generate_team_rosters(player_pool, n_teams=6) team_vals = [float(np.sum(r["projected_points"])) + np.random.normal(0, 5) for r in rosters] stf = SetTransformer(use_torch=False) stf.fit(rosters, team_vals) base_val = stf.predict(rosters[0]) top_a = player_pool.nlargest(3, "projected_points") new_p = top_a.head(1).copy() added = stf.value_added(rosters[0], new_p) redundancy = stf.get_redundancy_score(rosters[0]) results["set"] = {"base_value": float(base_val), "marginal_value": float(added), "redundancy": float(redundancy)} except Exception as e: results["set"] = {"error": str(e)} # 12 ─ Causal Forest try: from src.models.causal_forest import TransferCausalModel Xc = player_pool.head(300)[["projected_points", "games_played", "starter_pct", "goals_season", "assists_season"]].fillna(0).copy() Xc["role"] = player_pool.head(300)["role"] Xc["team_strength"] = np.random.uniform(0.5, 1.5, 300) Tc = pd.DataFrame({ "role": player_pool.head(300)["role"], "projected_points": player_pool.head(300)["projected_points"], "days_since_last_transfer": np.random.randint(1, 30, 300), }) Yc = pd.Series(np.random.randn(300) * 0.5 + player_pool.head(300)["projected_points"].values * 0.3) cf = TransferCausalModel() cf.fit(Xc, Tc, Yc) result_causal = cf.predict_effect(Xc.head(10), Tc.head(10)) results["causal"] = {"ate": float(result_causal.get("ate", 0)), "ate_lower": float(result_causal.get("ate_lower", -1)), "ate_upper": float(result_causal.get("ate_upper", 1))} except Exception as e: results["causal"] = {"error": str(e)} return results # ─── Charts ───────────────────────────────────────────────────────── def _chart_quantile(preds, risk, players): fig = make_subplots(rows=1, cols=2, subplot_titles=("Quantile Predictions (Top 30)", "Downside Risk P(FV < 5.5)")) n = min(30, len(players)) x = list(range(n)) fig.add_trace(go.Scatter(x=x, y=preds["P90"][:n], name="P90 (upside)", line=dict(color=PITCH_GREEN, width=2, dash="dot"), hovertext=players[:n]), row=1, col=1) fig.add_trace(go.Scatter(x=x, y=preds["P50"][:n], name="P50 (median)", line=dict(color=SKY, width=2.5), hovertext=players[:n]), row=1, col=1) fig.add_trace(go.Scatter(x=x, y=preds["P10"][:n], name="P10 (floor)", line=dict(color=RED, width=2, dash="dot"), fill="tonexty", fillcolor="rgba(255,77,94,0.08)"), row=1, col=1) fig.add_trace(go.Histogram(x=risk, nbinsx=25, name="P(Vote < 5.5)", marker_color=RED, opacity=0.7), row=1, col=2) fig.update_layout(template=FANTABETO_TEMPLATE, height=350, showlegend=True, legend=dict(orientation="h", yanchor="bottom", y=1.02)) fig.update_xaxes(title_text="Player (sorted by projection)", row=1, col=1) fig.update_yaxes(title_text="Fantavoto", row=1, col=1) fig.update_yaxes(title_text="Players", row=1, col=2) return fig def _chart_survival(expected, lower, upper, probs): fig = make_subplots(rows=1, cols=2, subplot_titles=("Minutes Distribution (95% CI)", "Starter Probability P(≥60 min)")) n = min(30, len(expected)) x = list(range(n)) fig.add_trace(go.Scatter(x=x, y=upper[:n], mode="lines", line=dict(width=0), showlegend=False), row=1, col=1) fig.add_trace(go.Scatter(x=x, y=lower[:n], mode="lines", fill="tonexty", fillcolor="rgba(56,189,248,0.15)", line=dict(width=0), name="95% CI"), row=1, col=1) fig.add_trace(go.Scatter(x=x, y=expected[:n], mode="lines+markers", line=dict(color=SKY, width=2.5), marker=dict(size=4, color=SKY), name="Expected min"), row=1, col=1) fig.add_trace(go.Scatter(x=x, y=[90]*n, mode="lines", line=dict(color=TEXT_SECONDARY, width=0.5, dash="dash"), name="Full 90"), row=1, col=1) fig.add_trace(go.Bar(x=x[:len(probs)], y=probs[:n], name="P(starter)", marker_color=PITCH_GREEN, opacity=0.8, text=[f"{p:.0%}" for p in probs[:n]], textposition="outside", textfont=dict(size=8)), row=1, col=2) fig.add_hline(y=0.7, line=dict(color=GOLD, width=1, dash="dot"), row=1, col=2) fig.update_layout(template=FANTABETO_TEMPLATE, height=350) fig.update_xaxes(title_text="Player", row=1, col=1) fig.update_yaxes(title_text="Minutes", range=[0, 95], row=1, col=1) fig.update_yaxes(title_text="Probability", range=[0, 1.05], row=1, col=2) return fig def _chart_bayesian(mean, std, reliability, players): fig = make_subplots(rows=1, cols=2, subplot_titles=("Bayesian Estimates ± Uncertainty", "Reliability Score (0-1)")) n = min(30, len(mean)) x = list(range(n)) fig.add_trace(go.Scatter( x=x, y=mean[:n], mode="markers", error_y=dict(type="data", array=std[:n], visible=True, color=VIOLET, thickness=1.5), marker=dict(size=7, color=VIOLET), name="Bayesian estimate", hovertext=players[:n], ), row=1, col=1) bar_colors = [PITCH_GREEN if r > 0.7 else (GOLD if r > 0.4 else RED) for r in reliability[:n]] fig.add_trace(go.Bar(x=x, y=reliability[:n], marker_color=bar_colors, opacity=0.85, name="Reliability", text=players[:n], hovertemplate="%{text}: %{y:.3f}"), row=1, col=2) fig.update_layout(template=FANTABETO_TEMPLATE, height=350) fig.update_xaxes(title_text="Player", row=1, col=1) fig.update_yaxes(title_text="Fantavoto", row=1, col=1) fig.update_yaxes(title_text="Score", range=[0, 1.05], row=1, col=2) return fig def _chart_conformal(preds, lowers, uppers, coverage): n = min(15, len(preds)) x = list(range(n)) fig = go.Figure() fig.add_trace(go.Scatter(x=x, y=uppers[:n], mode="lines", line=dict(width=0), showlegend=False)) fig.add_trace(go.Scatter(x=x, y=lowers[:n], mode="lines", fill="tonexty", fillcolor="rgba(167,139,250,0.15)", line=dict(width=0), name=f"{coverage*100:.0f}% band")) fig.add_trace(go.Scatter(x=x, y=preds[:n], mode="lines+markers", line=dict(color=VIOLET, width=2.5), marker=dict(size=5, color=VIOLET), name="Prediction")) fig.update_layout(template=FANTABETO_TEMPLATE, height=300) fig.update_xaxes(title_text="Player (holdout)") fig.update_yaxes(title_text="Fantavoto") return fig def _chart_bandit(trial_bids): fig = make_subplots(rows=1, cols=2, subplot_titles=("Bandit Bidding Over Time", "Bid Distribution by Role")) bids_by_role = {"P": [], "D": [], "C": [], "A": []} for b in trial_bids: bids_by_role.get(b["role"], []).append(b["bid"]) for role in ["P", "D", "C", "A"]: if bids_by_role[role]: y = bids_by_role[role] fig.add_trace(go.Scatter( x=list(range(len(y))), y=y, mode="lines+markers", name=f"{ROLE_ICONS[role]} {role}", line=dict(color=ROLE_COLORS.get(role, SKY), width=1.8), marker=dict(size=4, color=ROLE_COLORS.get(role, SKY)), ), row=1, col=1) for role in ["P", "D", "C", "A"]: if bids_by_role[role]: fig.add_trace(go.Box(y=bids_by_role[role], name=f"{role}", marker_color=ROLE_COLORS.get(role, SKY)), row=1, col=2) fig.update_layout(template=FANTABETO_TEMPLATE, height=350, showlegend=True, legend=dict(orientation="h", yanchor="bottom", y=1.02)) fig.update_xaxes(title_text="Bid #", row=1, col=1) fig.update_yaxes(title_text="Bid (cr)", row=1, col=1) return fig def _chart_opponent_bids(bids, players): if not bids: return go.Figure() n = min(30, len(bids), len(players)) fig = go.Figure(data=[go.Bar( x=players[:n], y=bids[:n], marker_color=SKY, opacity=0.8, text=[f"{b:.0f}" for b in bids[:n]], textposition="outside", )]) fig.update_layout(template=FANTABETO_TEMPLATE, height=280) fig.update_xaxes(title_text="Player", tickangle=-45, tickfont=dict(size=9)) fig.update_yaxes(title_text="Predicted Opponent Bid (cr)") return fig def _chart_budget_opt(allocation): labels = list(allocation.keys()) values = list(allocation.values()) colors_list = [ROLE_COLORS.get(r, SKY) for r in labels] fig = go.Figure(data=[go.Pie( labels=labels, values=values, hole=0.5, marker_colors=colors_list, textinfo="label+value", texttemplate="%{label}: %{value:.0f} cr", )]) fig.update_layout(template=FANTABETO_TEMPLATE, height=300) return fig def _chart_chemistry(bonus_matrix, labels): fig = go.Figure(data=go.Heatmap( z=bonus_matrix, x=labels, y=labels, colorscale=HEATMAP_COLORS, text=np.round(bonus_matrix, 3), texttemplate="%{text}", textfont=dict(size=9), zmin=0, zmax=1, )) fig.update_layout(template=FANTABETO_TEMPLATE, height=300, xaxis=dict(side="top"), yaxis=dict(autorange="reversed")) return fig def _chart_hawkes(statuses): labels = list(statuses.keys()) vals = list(statuses.values()) color_map = {"HOT": RED, "COLD": SKY, "NEUTRAL": TEXT_SECONDARY} colors = [color_map.get(v, TEXT_SECONDARY) for v in vals] fig = go.Figure(data=[go.Bar(x=[l.split()[-1] for l in labels], y=[1]*len(labels), marker_color=colors, text=vals, textposition="auto", textfont=dict(color=WHITE, size=11))]) fig.update_layout(template=FANTABETO_TEMPLATE, height=200, showlegend=False, yaxis=dict(showticklabels=False)) return fig # ─── Main page ────────────────────────────────────────────────────── def run(): inject_css() st.markdown("## 🔬 Dev Preview — 12 ML Models on Real 26/27 Serie A Data") st.caption("Models trained on 505 players, 2,000+ historical votes, 161 FBref features.") # Load data with st.spinner("Loading real Serie A data...", show_time=True): projections, stats, votes, roster, auction_plan = _load_real_data() player_pool = _build_rich_player_pool(projections, stats) with st.spinner("Building interaction graph & generating training data...", show_time=True): interaction_data = _generate_interaction_data(player_pool) auction_logs = _generate_auction_logs(player_pool) with st.spinner("Training 12 ML models (this takes ~10s)...", show_time=True): results = _init_models(player_pool, interaction_data, auction_logs) st.divider() # ── Header KPIs ── k1, k2, k3, k4, k5, k6 = st.columns(6) phases_ok = sum(1 for v in results.values() if isinstance(v, dict) and "error" not in v) with k1: st.markdown(kpi_card("MODELS ACTIVE", f"{phases_ok}/12", "12 ML models", PITCH_GREEN), unsafe_allow_html=True) with k2: st.markdown(kpi_card("PLAYERS", str(len(player_pool)), "Serie A 26/27", SKY), unsafe_allow_html=True) with k3: top_fv = player_pool["projected_points"].max() top_name = player_pool.loc[player_pool["projected_points"].idxmax(), "name"] st.markdown(kpi_card("TOP PROJ FV", f"{top_fv:.2f}", top_name, GOLD), unsafe_allow_html=True) with k4: st.markdown(kpi_card("FEATURES", str(161), "FBref + Fantacalcio", VIOLET), unsafe_allow_html=True) with k5: st.markdown(kpi_card("HISTORICAL VOTES", f"{len(votes):,}", "matchday records", PITCH_GREEN), unsafe_allow_html=True) with k6: cf_ate = results.get("causal", {}).get("ate", 0) st.markdown(kpi_card("CAUSAL ATE", f"{cf_ate:+.3f}", "transfer effect", GOLD), unsafe_allow_html=True) st.divider() # ── Live Auction Simulator ── st.markdown("### 🎮 Live Auction Simulator") st.caption("Real players, real projections. Simulate a bidding round with bandit + opponent models.") col_sim1, col_sim2, col_sim3 = st.columns(3) with col_sim1: role_filter = st.selectbox("Filter by Role", ["All", "P", "D", "C", "A"]) with col_sim2: sim_budget = st.slider("Your Budget (cr)", 50, 500, 400, step=10) with col_sim3: filtered_pool = player_pool if role_filter == "All" else player_pool[player_pool["role"] == role_filter] top_players = filtered_pool.nlargest(100, "projected_points")["name"].tolist() sim_player = st.selectbox("Target Player", top_players[:50] if len(top_players) > 50 else top_players) if st.button("🎯 Simulate Bid Decision", type="primary"): try: from src.optimization.bandit_auction import BanditAuctionSolver from src.optimization.auction_solver import AuctionConfig, PlayerValuation from src.optimization.opponent_bidding_model import OpponentBidModel from src.optimization.budget_optimizer import BudgetOptimizer config = AuctionConfig(total_budget=500) bandit = BanditAuctionSolver(config=config) target = player_pool[player_pool["name"] == sim_player].iloc[0] qe_result = results.get("quantile", {}) if "preds" in qe_result: try: idx = qe_result["players"].index(sim_player) p10 = qe_result["preds"]["P10"][idx] p50 = qe_result["preds"]["P50"][idx] p90 = qe_result["preds"]["P90"][idx] except (ValueError, IndexError, KeyError): pt = target["projected_points"] p10, p50, p90 = pt * 0.85, pt, pt * 1.15 else: pt = target["projected_points"] p10, p50, p90 = pt * 0.85, pt, pt * 1.15 pv = PlayerValuation( name=target["name"], team=target["team"], role=target["role"], projected_points=float(target["projected_points"]), market_value=float(target["market_value"]), ceiling_price=float(target["projected_points"] * 5), ) state = { "budget_remaining": sim_budget, "total_budget": 500, "slots_remaining": {"P": 1, "D": 3, "C": 4, "A": 2}, "role_quotas": {"P": 3, "D": 8, "C": 8, "A": 6}, "slot_quotas": {"P": 3, "D": 8, "C": 8, "A": 6}, "round_number": 4, "total_rounds": 20, "opponent_budgets": [sim_budget - 20, sim_budget + 10, sim_budget - 50], "players_remaining_in_role": {"P": 15, "D": 50, "C": 50, "A": 30}, "player_pool": [pv], } arm_idx, bandit_bid = bandit.select_bid(pv, state) obm = OpponentBidModel() bid_row = pd.DataFrame([{ "player_name": target["name"], "player_role": target["role"], "player_projected_points": target["projected_points"], }]) opp_state_full = { "budget_remaining": sim_budget, "total_budget": 500, "initial_budget": 500, "slots_total": {"P": 3, "D": 8, "C": 8, "A": 6}, "slots_filled": {"P": 1, "D": 2, "C": 2, "A": 2}, "slots_remaining": {"P": 2, "D": 6, "C": 6, "A": 4}, "role_quotas": {"P": 3, "D": 8, "C": 8, "A": 6}, "aggression_factor": 1.0, } opp_bid = float(obm.predict_opponent_bids(bid_row, opp_state_full).values[0]) bo = BudgetOptimizer(total_budget=500) role_budget = bo.optimize(player_pool.head(100), n_calls=5) role_rec = role_budget.get(target["role"], 50) win_pct = bandit_bid / max(bandit_bid + opp_bid, 1) * 100 risk_label = "LOW" if p10 > 5.5 else ("MED" if p10 > 4.5 else "HIGH") risk_color = PITCH_GREEN if risk_label == "LOW" else (GOLD if risk_label == "MED" else RED) with st.container(): st.markdown("### 📊 Bid Decision") kd1, kd2, kd3, kd4, kd5 = st.columns(5) with kd1: bid_color = PITCH_GREEN if bandit_bid > opp_bid else RED st.markdown(kpi_card("ML BID", f"{bandit_bid} cr", f"Opp ~{opp_bid:.0f} cr", bid_color), unsafe_allow_html=True) with kd2: st.markdown(kpi_card("WIN PROB", f"{min(win_pct, 95):.0f}%", "", GOLD), unsafe_allow_html=True) with kd3: st.markdown(kpi_card("PROJECTION", f"{p50:.2f}", f"P10:{p10:.1f} P90:{p90:.1f}", SKY), unsafe_allow_html=True) with kd4: st.markdown(kpi_card("ROLE BUDGET", f"{role_rec:.0f} cr", f"{target['role']} allocation", VIOLET), unsafe_allow_html=True) with kd5: st.markdown(kpi_card("RISK", risk_label, f"VaR floor: {p10:.1f}", risk_color), unsafe_allow_html=True) value_ratio = target["projected_points"] / max(bandit_bid, 1) if bandit_bid > opp_bid: insight(f"🟢 **BUY** — Bandit recommends {bandit_bid}cr for **{target['name']}** " f"({ROLE_ICONS.get(target['role'], '')} {target['team']}, FV {target['projected_points']:.2f}). " f"Expected to beat opponent ~{opp_bid:.0f}cr. Value: {value_ratio:.3f} FV/cr.") else: insight(f"🔴 **PASS** — Opponent likely bids higher (~{opp_bid:.0f}cr). " f"Consider a budget reallocation or targeting an alternative in the {target['role']} role.") except Exception as e: st.error(f"Simulation error: {e}") st.divider() # ── Phase tabs ── tab1, tab2, tab3, tab4, tab5, tab6 = st.tabs([ "⚡ Phase 1: Quantile & Survival", "🎰 Phase 2: Adaptive Auction", "🧠 Phase 3: RL & Set Transformer", "📊 Phase 4: Bayesian & Conformal", "🔗 Phase 5: Chemistry & Form", "🔬 Phase 6: Causal Inference", ]) with tab1: section("⚡ Quantile Ensemble — Risk-Aware Projections") q = results.get("quantile", {}) if "error" in q: st.warning(q["error"]) else: fig = _chart_quantile(q["preds"], q["risk"], q.get("players", [])) st.plotly_chart(fig, width="stretch") insight("P10/P50/P90 predictions on top 60 Serie A players. " "The downside risk histogram shows how many players risk scoring below 5.5 — " "critical for avoiding zero-score matchdays.") section("⏱ Minutes Survival Model — Playing Time Distribution") s = results.get("survival", {}) if "error" in s: st.warning(s["error"]) else: fig = _chart_survival(s["expected"], s["lower"], s["upper"], s["starter_probs"]) st.plotly_chart(fig, width="stretch") insight("Weibull AFT trained on real minutes data. Wide confidence bands on players " "with irregular playing patterns. Starter probability < 70% = auction risk flag.") with tab2: section("🎰 Thompson Sampling — Live Auction Strategy") b = results.get("bandit", {}) if "error" in b: st.warning(b["error"]) else: fig = _chart_bandit(b["trial_bids"]) st.plotly_chart(fig, width="stretch") insight("Bandit learns bid amounts interactively. Forwards get higher bids; " "exploration bonus encourages discovering undervalued players early in auction.") section("💰 Opponent Bidding Model") ob = results.get("opponent_bidding", {}) if "error" in ob: st.warning(ob["error"]) else: fig = _chart_opponent_bids(ob["sample_bids"], ob.get("players", [])) st.plotly_chart(fig, width="stretch") insight("LightGBM predicts competitor max bids from role, scarcity, and player quality. " "Don't overpay when nobody wants the player.") section("📐 Bayesian Budget Optimization") bo_r = results.get("budget_opt", {}) if "error" in bo_r: st.warning(bo_r["error"]) else: fig = _chart_budget_opt(bo_r["allocation"]) st.plotly_chart(fig, width="stretch") insight("GP optimization finds optimal budget split across GK/DEF/MID/FWD roles " "based on the real player pool's value distribution. Adapts to market quality.") with tab3: section("🧠 Double DQN Auction Agent") rl = results.get("rl", {}) if "error" in rl: st.warning(rl["error"]) else: rk1, rk2, rk3 = st.columns(3) with rk1: st.markdown(kpi_card("STATE DIM", str(rl["observation_dim"]), "features", SKY), unsafe_allow_html=True) with rk2: st.markdown(kpi_card("ACTIONS", str(rl["action_dim"]), "bid levels", GOLD), unsafe_allow_html=True) with rk3: st.markdown(kpi_card("REPLAY", str(rl["buffer_size"]), "experiences", VIOLET), unsafe_allow_html=True) insight("RL agent trained on 30 simulated auctions. Learns to hold budget for later " "rounds when high-value players typically appear. 5000+ episode training recommended.") section("🧩 Set Transformer — Team Composition Value") sf = results.get("set", {}) if "error" in sf: st.warning(sf["error"]) else: sk1, sk2, sk3 = st.columns(3) with sk1: st.markdown(kpi_card("TEAM VALUE", f"{sf['base_value']:.0f} PTS", "set-based estimate", SKY), unsafe_allow_html=True) with sk2: dc = PITCH_GREEN if sf["marginal_value"] > 0 else RED st.markdown(kpi_card("MARGINAL Δ", f"{sf['marginal_value']:+.1f}", "add 1 player", dc), unsafe_allow_html=True) with sk3: rc = PITCH_GREEN if sf["redundancy"] < 0.5 else GOLD st.markdown(kpi_card("REDUNDANCY", f"{sf['redundancy']:.2f}", "0=diverse 1=overlap", rc), unsafe_allow_html=True) with tab4: section("📊 Bayesian Hierarchical Pooling") bp = results.get("bayesian", {}) if "error" in bp: st.warning(bp["error"]) else: fig = _chart_bayesian(bp["mean"], bp["std"], bp["reliability"], bp.get("players", [])) st.plotly_chart(fig, width="stretch") insight("Players with < 10 matches get heavy shrinkage toward role mean. " "Low reliability = don't pay premium for unproven talent.") section("🎯 Conformal Prediction — Calibrated Bands") cp_r = results.get("conformal", {}) if "error" in cp_r: st.warning(cp_r["error"]) else: fig = _chart_conformal(cp_r["predictions"], cp_r["lowers"], cp_r["uppers"], cp_r["coverage"]) st.plotly_chart(fig, width="stretch") insight(f"Conformal coverage: {cp_r['coverage']*100:.0f}% on 20 holdout players. " "Model-agnostic, no distributional assumptions needed — just works with any underlying predictor.") with tab5: section("🔗 Graph Attention Network — Player Chemistry") ch = results.get("chemistry", {}) if "error" in ch: st.warning(ch["error"]) else: fig = _chart_chemistry(ch["bonus_matrix"], ch.get("labels", [])) st.plotly_chart(fig, width="stretch") insight("Pairwise chemistry from synthetic pass/assist/cross networks. Higher values = " "stronger link between players on the same team. Target linked pairs in auction.") section("🔥 Hawkes Process — Form Momentum") hf_r = results.get("hawkes", {}) if "error" in hf_r: st.warning(hf_r["error"]) else: fig = _chart_hawkes(hf_r["statuses"]) st.plotly_chart(fig, width="stretch") insight("HOT = positive momentum (buy window open), COLD = negative drift (wait), " "NEUTRAL = baseline. Self-exciting process captures temporary scoring bursts.") with tab6: section("🔬 Causal Forest — Transfer Effects") cf = results.get("causal", {}) if "error" in cf: st.warning(cf["error"]) else: ck1, ck2, ck3 = st.columns(3) with ck1: c = PITCH_GREEN if cf["ate"] > 0 else RED st.markdown(kpi_card("ATE", f"{cf['ate']:+.4f}", "avg tx effect", c), unsafe_allow_html=True) with ck2: st.markdown(kpi_card("CI LOWER", f"{cf['ate_lower']:+.4f}", "95%", TEXT_SECONDARY), unsafe_allow_html=True) with ck3: st.markdown(kpi_card("CI UPPER", f"{cf['ate_upper']:+.4f}", "95%", TEXT_SECONDARY), unsafe_allow_html=True) insight("Causal inference separates true player impact from confounding (team, schedule, luck). " "Adding a highly-projected player doesn't always improve team score when controlling for roster fit.") # ── Auction Plan reference ── st.divider() section("📋 Reference: Existing Al-Cihred Auction Plan (25 players)") st.dataframe( auction_plan[["player", "team", "role", "bid_cap", "fv_proj", "games", "starter%"]] .rename(columns={"starter%": "start_pct"}) .style.background_gradient(subset=["fv_proj", "bid_cap"], cmap="viridis"), height=600, width="stretch", ) insight("25-player squad from the MILP solver. Use new ML models above to refine bids and alternatives.") st.divider() with st.expander("⚙️ Methodology — 12 Models on Real Data"): st.markdown(""" | # | Model | Data Used | |---|-------|----------| | 1 | QuantileEnsemble | 505 player projections, 161 FBref features | | 2 | MinutesSurvivalModel | Real minutes played, starter percentages | | 3 | BanditAuctionSolver | Real QI/QA/FVM prices, projected FV | | 4 | OpponentBidModel | QI-based pricing, role scarcity from roster | | 5 | BudgetOptimizer | Real pool value distribution | | 6 | PlayerChemistryGAT | 800 synthetic pass/assist/cross edges | | 7 | PlayerFormModel | 2,021 historical matchday votes | | 8 | BayesianPlayerModel | Per-role shrinkage on 200 players | | 9 | ConformalPredictor | 300-player holdout calibration | | 10 | SetTransformer | 6 synthetic teams of 25 real players each | | 11 | RLAuctionPolicy | 30-episode training on real player pool | | 12 | TransferCausalModel | 300-player causal forest | """, unsafe_allow_html=False) st.divider() st.caption(f"Real data: {len(player_pool)} Serie A 26/27 players · {len(votes):,} historical votes · " f"FBref stats · Fantacalcio.it QI/QA/FVM · Project Al-Cihred auction plan") if __name__ == "__main__": run()