feat: add dev preview dashboard showcasing all 12 new ML models

- New page 06_dev_preview.py: interactive ML model showcase
- All 12 models initialized from synthetic data with live charts
- Live Auction Simulator: bandit + opponent model + budget optimizer
- Tabbed UI: 6 tabs, one per phase
- Resilient warehouse: returns typed empty DataFrames when no data
- Dev Preview set as default landing page for demo mode
- Live at http://localhost:8507

dashboard/app.py

@@ -19,6 +19,7 @@ from dashboard.viz.components import inject_css
 from dashboard.viz.template import stub_render  # triggers template registration
 
 PAGES = {
+    "🔬 Dev Preview": "dashboard.pages.06_dev_preview",
     "⚽ Matchday": "dashboard.pages.01_matchday",
     "👤 Players": "dashboard.pages.02_players",
     "💰 Auction": "dashboard.pages.03_auction",

dashboard/pages/01_matchday.py

@@ -69,6 +69,9 @@ def _build_fixture_heatmap(players, fixtures):
 def run():
     inject_css()
     preds, fixtures, players, lineups, votes = _get_data()
+    if preds.empty or len(preds) <= 1:
+        st.warning("No warehouse data. Run pipeline or use Dev Preview tab.")
+        return
 
     projected, league_avg, risk_count, trend = _compute_kpis(preds, players, votes)
 

dashboard/pages/06_dev_preview.py

@@ -0,0 +1,875 @@
+"""Page 6 — Dev Preview: New ML Models for Auction Optimization.
+
+Showcases all 10 new ML modules from Phases 1-6 with interactive visualizations.
+Runs on synthetic data so it works without the full data pipeline.
+"""
+
+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,
+)
+
+
+# ─── Synthetic data generation ──────────────────────────────────────
+
+def _generate_player_pool(n_players=120, seed=42):
+    rng = np.random.RandomState(seed)
+    roles_dist = ["P"] * 10 + ["D"] * 40 + ["C"] * 40 + ["A"] * 30
+    teams = [f"Team_{i}" for i in range(20)]
+    data = []
+    for i in range(n_players):
+        role = roles_dist[i % len(roles_dist)]
+        base_fv = {"P": 6.2, "D": 6.3, "C": 6.5, "A": 7.0}[role]
+        fv = base_fv + rng.normal(0, 0.8)
+        fv = max(4.5, min(9.5, fv))
+        qi = np.exp(fv - 3.5) * rng.uniform(0.8, 1.5)
+        games = int(rng.choice([38, 35, 30, 25, 20, 15, 10, 5], p=[0.15, 0.15, 0.2, 0.15, 0.12, 0.1, 0.08, 0.05]))
+        minutes_last_3 = rng.uniform(0, 90) if rng.random() < 0.7 else rng.uniform(50, 90)
+        data.append({
+            "name": f"Player_{i}",
+            "role": role,
+            "team": rng.choice(teams),
+            "projected_points": round(fv, 2),
+            "fv_std": round(rng.uniform(0.3, 1.5), 2),
+            "market_value": round(qi, 1),
+            "starter_pct": round(rng.uniform(0.3, 0.98), 2),
+            "minutes_last_3": round(minutes_last_3, 1),
+            "games_played": games,
+            "xg_p90": round(rng.uniform(0.01, 0.8), 3),
+            "xa_p90": round(rng.uniform(0.01, 0.5), 3),
+            "goals_season": round(rng.poisson(max(fv - 5.5, 0.01)) if fv > 5.5 else 0),
+            "assists_season": round(rng.poisson(max((fv - 5.5) * 0.5, 0.01)) if fv > 5.5 else 0),
+            "yellow_per_game": round(rng.beta(2, 20), 3),
+            "red_per_game": round(rng.beta(1, 50), 4),
+            "rest_days": int(rng.uniform(2, 10)),
+            "fatigue_rolling_3": round(rng.uniform(0, 90), 1),
+            "feature1": round(rng.normal(0, 1), 2),
+            "feature2": round(rng.normal(0, 1), 2),
+        })
+    return pd.DataFrame(data)
+
+
+def _generate_team_rosters(player_pool, n_teams=10, seed=42):
+    rng = np.random.RandomState(seed)
+    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
+
+
+def _generate_interaction_data(player_pool, seed=42):
+    rng = np.random.RandomState(seed)
+    rows = []
+    players = player_pool["name"].tolist()
+    for _ in range(300):
+        a = rng.choice(players)
+        b = rng.choice(players)
+        if a == b:
+            continue
+        rows.append({
+            "player": a,
+            "teammate": b,
+            "passes_to": rng.randint(0, 15),
+            "assists_to": rng.randint(0, 2),
+            "crosses_to": rng.randint(0, 5),
+            "matchday": rng.randint(1, 39),
+        })
+    return pd.DataFrame(rows)
+
+
+def _generate_auction_logs(player_pool, seed=42):
+    rng = np.random.RandomState(seed)
+    rows = []
+    for _ in range(500):
+        player = player_pool.iloc[rng.randint(0, len(player_pool))]
+        rows.append({
+            "player_name": player["name"],
+            "player_role": player["role"],
+            "player_projected_points": player["projected_points"],
+            "opponent_budget_remaining": rng.uniform(50, 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(player["market_value"] * rng.uniform(0.5, 2.0))),
+        })
+    return pd.DataFrame(rows)
+
+
+# ─── Model initialization cache ─────────────────────────────────────
+
+@st.cache_resource
+def _init_models(player_pool, interaction_data, auction_logs):
+    results = {}
+
+    # Phase 1: Quantile Ensemble
+    try:
+        from src.models.quantile_model import QuantileEnsemble
+        X = player_pool[["projected_points", "fv_std", "minutes_last_3", "games_played",
+                          "xg_p90", "xa_p90", "rest_days", "fatigue_rolling_3",
+                          "feature1", "feature2"]].fillna(0)
+        y = player_pool["projected_points"]
+        qe = QuantileEnsemble(quantiles=(0.10, 0.50, 0.90), n_estimators=50)
+        qe.fit(X, y)
+        preds = qe.predict(X.head(60))
+        risk = qe.predict_downside_risk(X.head(60), threshold=5.5)
+        results["quantile"] = {"model": qe, "preds": preds, "risk": risk, "X": X.head(60)}
+    except Exception as e:
+        results["quantile"] = {"error": str(e)}
+
+    # Phase 1: Survival Model
+    try:
+        from src.models.survival_model import MinutesSurvivalModel
+        surv_df = player_pool[
+            ["minutes_last_3", "games_played", "rest_days", "fatigue_rolling_3",
+             "feature1", "feature2"]
+        ].fillna(0)
+        durations = np.clip(player_pool["minutes_last_3"].values + np.random.normal(0, 10, len(player_pool)), 1, 90)
+        events = (player_pool["starter_pct"].values > 0.5).astype(int)
+        ms = MinutesSurvivalModel(force_scipy=True)
+        ms.fit(surv_df, durations, events)
+        expected, lower, upper = ms.predict_distribution(surv_df.head(30))
+        starter_probs = ms.predict_starter_probability(surv_df.head(30))
+        results["survival"] = {"model": ms, "expected": expected.tolist(),
+                                "lower": lower.tolist(), "upper": upper.tolist(),
+                                "starter_probs": starter_probs.tolist()}
+    except Exception as e:
+        results["survival"] = {"error": str(e)}
+
+    # Phase 4: Bayesian Pooling
+    try:
+        from src.models.bayesian_pooling import BayesianPlayerModel
+        Xb = player_pool[["role", "projected_points", "minutes_last_3", "games_played"]].copy()
+        yb = player_pool["projected_points"]
+        bp = BayesianPlayerModel()
+        bp.fit(Xb, yb)
+        mean, std = bp.predict_with_uncertainty(Xb.head(30))
+        reliability = bp.get_player_reliability(Xb.head(30))
+        results["bayesian"] = {"model": bp, "mean": mean.tolist(),
+                                "std": std.tolist(), "reliability": reliability.tolist()}
+    except Exception as e:
+        results["bayesian"] = {"error": str(e)}
+
+    # Phase 2: Bandit Auction
+    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 _ in range(20):
+            player = player_pool.iloc[np.random.randint(0, 60)]
+            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": 500 - _ * 20,
+                "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": _ + 1,
+                "total_rounds": 20,
+                "opponent_budgets": [400, 350, 420],
+                "players_remaining_in_role": {"P": 5, "D": 10, "C": 10, "A": 8},
+                "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)}
+
+    # Phase 5: GAT Chemistry
+    try:
+        from src.models.gat_model import PlayerChemistryGAT
+        gat = PlayerChemistryGAT()
+        gat.build_graph(interaction_data)
+        chem_features = gat.extract_interaction_features("Player_0", player_pool.head(5)["name"].tolist())
+        bonus_matrix = []
+        for p in ["Player_0", "Player_1", "Player_2", "Player_3", "Player_4"]:
+            row = []
+            for q in ["Player_0", "Player_1", "Player_2", "Player_3", "Player_4"]:
+                row.append(gat.compute_interaction_bonus(p, q))
+            bonus_matrix.append(row)
+        results["chemistry"] = {"features": chem_features, "bonus_matrix": bonus_matrix}
+    except Exception as e:
+        results["chemistry"] = {"error": str(e)}
+
+    # Phase 5: Hawkes Form
+    try:
+        from src.models.hawkes_form import PlayerFormModel
+        dates = pd.date_range("2023-08-20", periods=38, freq="7D")
+        form_data = pd.DataFrame({
+            "player": np.repeat(["Player_0", "Player_1", "Player_5", "Player_10", "Player_20"], 8)[:38][:30],
+            "match_date": dates[:30],
+            "minutes": np.random.uniform(30, 90, 30),
+        })
+        form_y = pd.Series(np.random.randn(30) * 1.5 + 6.5)
+        hf = PlayerFormModel()
+        hf.fit(form_data, form_y)
+        statuses = hf.get_form_status(form_data)
+        results["hawkes"] = {"statuses": dict(zip(form_data["player"].tolist()[:10], statuses[:10]))}
+    except Exception as e:
+        results["hawkes"] = {"error": str(e)}
+
+    # Phase 3: RL Auction
+    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(30), n_opponents=2, config=config_small)
+        agent = RLAuctionPolicy(state_dim=14, action_dim=11, hidden_dim=32)
+        trainer = RLAuctionTrainer(
+            player_pool.head(30), n_opponents=2, config=config_small,
+        )
+        agent = trainer.train(n_episodes=20, 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)}
+
+    # Phase 3: Set Transformer
+    try:
+        from src.models.set_transformer import SetTransformer
+        rosters = _generate_team_rosters(player_pool, n_teams=6)
+        team_vals = [
+            sum(r["projected_points"]) + np.random.normal(0, 8)
+            for r in rosters
+        ]
+        stf = SetTransformer(use_torch=False)
+        stf.fit(rosters, team_vals)
+        base_val = stf.predict(rosters[0])
+        new_p = pd.DataFrame([{
+            "name": "NewPlayer", "role": "A", "feature1": 1.5,
+            "projected_points": 8.5, "team": "Team_0",
+        }])
+        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)}
+
+    # Phase 6: Causal Forest
+    try:
+        from src.models.causal_forest import TransferCausalModel
+        Xc = player_pool.head(200)[["projected_points", "minutes_last_3", "games_played",
+                                      "feature1", "feature2"]].copy()
+        Xc["role"] = player_pool.head(200)["role"]
+        Xc["team_strength"] = np.random.uniform(0.5, 1.5, 200)
+        Tc = pd.DataFrame({
+            "role": player_pool.head(200)["role"],
+            "projected_points": player_pool.head(200)["projected_points"],
+            "days_since_last_transfer": np.random.randint(1, 30, 200),
+        })
+        Yc = pd.Series(np.random.randn(200) + 6.5)
+        cf = TransferCausalModel()
+        cf.fit(Xc, Tc, Yc)
+        result_causal = cf.predict_effect(Xc.head(5), Tc.head(5))
+        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)}
+
+    # Phase 4: Conformal Predictor
+    try:
+        from src.models.conformal_predictor import ConformalPredictor
+        from sklearn.linear_model import Ridge
+        Xcp = player_pool[["minutes_last_3", "games_played", "xg_p90", "xa_p90",
+                            "fatigue_rolling_3", "feature1", "feature2"]].fillna(0)
+        ycp = player_pool["projected_points"]
+        base = Ridge(alpha=1.0)
+        base.fit(Xcp.head(100), ycp.head(100))
+        cp = ConformalPredictor(base, alpha=0.10)
+        cp.calibrate(Xcp.iloc[100:150], ycp.iloc[100:150])
+        yp, yl, yu = cp.predict_with_band(Xcp.iloc[150:160])
+        coverage = cp.coverage(Xcp.iloc[150:160], ycp.iloc[150:160])
+        results["conformal"] = {"coverage": float(coverage), "n_test": 10,
+                                 "predictions": yp.tolist()[:10],
+                                 "lowers": yl.tolist()[:10],
+                                 "uppers": yu.tolist()[:10]}
+    except Exception as e:
+        results["conformal"] = {"error": str(e)}
+
+    # Phase 2: Opponent Bidding
+    try:
+        from src.optimization.opponent_bidding_model import OpponentBidModel
+        obm = OpponentBidModel()
+        sample_players = player_pool.head(30).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)
+        my_bids_sample = np.array([10, 15, 20, 5, 8, 12, 3, 25, 18, 7])
+        if len(my_bids_sample) == min(10, len(bids)):
+            probs = obm.predict_p_acquire(sample_players.head(min(10, len(bids))), my_bids_sample[:min(10, len(bids))], opp_state)
+            results["opponent_bidding"] = {"sample_bids": bids.head(10).tolist() if len(bids) >= 10 else bids.tolist(),
+                                           "acq_probs": probs.tolist()[:10] if len(bids) >= 10 else []}
+        else:
+            results["opponent_bidding"] = {"sample_bids": bids.head(10).tolist()}
+    except Exception as e:
+        results["opponent_bidding"] = {"error": str(e)}
+
+    # Phase 2: Budget Optimizer
+    try:
+        from src.optimization.budget_optimizer import BudgetOptimizer
+        bo = BudgetOptimizer(total_budget=500)
+        allocation = bo.optimize(player_pool, n_calls=10)
+        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)}
+
+    return results
+
+
+# ─── Chart helpers ──────────────────────────────────────────────────
+
+def _phase_chart_quantile(preds, risk):
+    fig = make_subplots(rows=1, cols=2, subplot_titles=("Quantile Predictions", "Downside Risk Distribution"))
+    n = 30
+    x = list(range(n))
+    fig.add_trace(go.Scatter(x=x, y=preds["P90"].tolist()[:n], name="P90 (upside)",
+                               line=dict(color=PITCH_GREEN, width=2, dash="dot")), row=1, col=1)
+    fig.add_trace(go.Scatter(x=x, y=preds["P50"].tolist()[:n], name="P50 (median)",
+                               line=dict(color=SKY, width=2.5)), row=1, col=1)
+    fig.add_trace(go.Scatter(x=x, y=preds["P10"].tolist()[: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.tolist(), nbinsx=20, name="P(FV < 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", row=1, col=1)
+    fig.update_yaxes(title_text="FV", row=1, col=1)
+    fig.update_xaxes(title_text="P(downside)", row=1, col=2)
+    fig.update_yaxes(title_text="Count", row=1, col=2)
+    return fig
+
+
+def _phase_chart_survival(expected, lower, upper, probs):
+    fig = make_subplots(rows=1, cols=2, subplot_titles=("Minutes Distribution (P95)", "Starter Probability"))
+    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=5, color=SKY),
+                               name="Expected min"), row=1, col=1)
+    max_line = [90] * n
+    fig.add_trace(go.Scatter(x=x, y=max_line, mode="lines",
+                               line=dict(color=TEXT_SECONDARY, width=0.5, dash="dash"),
+                               name="Full match"), 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), 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_xaxes(title_text="Player", row=1, col=2)
+    fig.update_yaxes(title_text="P(≥60 min)", range=[0, 1], row=1, col=2)
+    return fig
+
+
+def _phase_chart_bayesian(mean, std, reliability):
+    fig = make_subplots(rows=1, cols=2, subplot_titles=("Predictions ± Uncertainty", "Reliability Score"))
+    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),
+        marker=dict(size=7, color=VIOLET),
+        name="Bayesian estimate",
+    ), row=1, col=1)
+    fig.add_trace(go.Bar(x=x, y=reliability[:n], marker_color=GOLD, opacity=0.8,
+                           name="Reliability"), 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="FV", row=1, col=1)
+    fig.update_xaxes(title_text="Player", row=1, col=2)
+    fig.update_yaxes(title_text="Score (0-1)", range=[0, 1], row=1, col=2)
+    return fig
+
+
+def _phase_chart_conformal(preds, lowers, uppers, coverage):
+    n = min(10, 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=6, color=VIOLET),
+                               name="Prediction"))
+    fig.update_layout(template=FANTABETO_TEMPLATE, height=300)
+    fig.update_xaxes(title_text="Player")
+    fig.update_yaxes(title_text="FV")
+    return fig
+
+
+def _phase_chart_bandit(trial_bids):
+    fig = make_subplots(rows=1, cols=2, subplot_titles=("Bid History", "Bid by Role"))
+    roles = ["P", "D", "C", "A"]
+    bids_by_role = {r: [] for r in roles}
+    for b in trial_bids:
+        bids_by_role.get(b["role"], []).append(b["bid"])
+    for role in roles:
+        if bids_by_role[role]:
+            y = bids_by_role[role]
+            x = list(range(len(y)))
+            fig.add_trace(go.Scatter(
+                x=x, y=y, mode="lines+markers", name=f"{ROLE_ICONS[role]} {role}",
+                line=dict(color=ROLE_COLORS.get(role, SKY), width=2),
+                marker=dict(size=6, color=ROLE_COLORS.get(role, SKY)),
+            ), row=1, col=1)
+    for role in roles:
+        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="Decision #", row=1, col=1)
+    fig.update_yaxes(title_text="Bid (cr)", row=1, col=1)
+    fig.update_yaxes(title_text="Bid (cr)", row=1, col=2)
+    return fig
+
+
+def _phase_chart_chemistry(bonus_matrix):
+    labels = ["P0", "P1", "P2", "P3", "P4"]
+    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=10),
+        zmin=0, zmax=1,
+    ))
+    fig.update_layout(template=FANTABETO_TEMPLATE, height=300,
+                       xaxis=dict(side="top"), yaxis=dict(autorange="reversed"))
+    return fig
+
+
+def _phase_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=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
+
+
+def _phase_chart_opponent_bids(bids):
+    if not bids:
+        return go.Figure()
+    fig = go.Figure(data=[go.Bar(
+        x=list(range(len(bids))), y=bids,
+        marker_color=SKY, opacity=0.8,
+        text=[f"{b:.0f}" for b in bids], textposition="outside",
+    )])
+    fig.update_layout(template=FANTABETO_TEMPLATE, height=250)
+    fig.update_xaxes(title_text="Player")
+    fig.update_yaxes(title_text="Predicted Opponent Bid (cr)")
+    return fig
+
+
+def _phase_chart_budget_opt(allocation):
+    labels = list(allocation.keys())
+    values = list(allocation.values())
+    colors = [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, textinfo="label+value",
+        texttemplate="%{label}: %{value:.0f} cr",
+    )])
+    fig.update_layout(template=FANTABETO_TEMPLATE, height=280)
+    return fig
+
+
+# ─── Main page ──────────────────────────────────────────────────────
+
+def run():
+    inject_css()
+    st.markdown("## 🔬 Dev Preview — New ML Models for Auction Optimization")
+    st.caption("Phase 1–6 models running on synthetic data. Interact with the auction components below.")
+
+    # Generate data
+    with st.spinner("Generating synthetic data & training models..."):
+        player_pool = _generate_player_pool(n_players=120)
+        interaction_data = _generate_interaction_data(player_pool)
+        auction_logs = _generate_auction_logs(player_pool)
+        results = _init_models(player_pool, interaction_data, auction_logs)
+
+    st.divider()
+
+    # ── Header KPIs ──
+    k1, k2, k3, k4, k5, k6 = st.columns(6)
+    phases_working = 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_working}/12", "", PITCH_GREEN), unsafe_allow_html=True)
+    with k2:
+        st.markdown(kpi_card("PLAYERS", str(len(player_pool)), "synthetic", SKY), unsafe_allow_html=True)
+    with k3:
+        st.markdown(kpi_card("AUCTION BUDGET", "500 cr", "total", GOLD), unsafe_allow_html=True)
+    with k4:
+        rl_size = results.get("rl", {}).get("buffer_size", 0)
+        st.markdown(kpi_card("RL BUFFER", str(rl_size), "experiences", VIOLET), unsafe_allow_html=True)
+    with k5:
+        st.markdown(kpi_card("INTERACTIONS", str(len(interaction_data)), "edges", PITCH_GREEN), unsafe_allow_html=True)
+    with k6:
+        cf_ate = results.get("causal", {}).get("ate", 0)
+        st.markdown(kpi_card("AVG CAUSAL EFFECT", f"{cf_ate:+.2f}", "ATE", GOLD), unsafe_allow_html=True)
+
+    st.divider()
+
+    # ── Live Auction Simulator ──
+    st.markdown("### 🎮 Live Auction Simulator")
+    st.caption("Run a mock auction round to see the bandit + opponent model + budget optimizer in action.")
+
+    c_sim1, c_sim2 = st.columns(2)
+    with c_sim1:
+        sim_budget = st.slider("Your Budget", 100, 700, 450, step=10)
+        sim_player = st.selectbox("Available Player for Bidding", player_pool.head(30)["name"].tolist())
+    with c_sim2:
+        sim_round = st.slider("Round", 1, 20, 5)
+        st.markdown(f"<br><span style='color:{TEXT_SECONDARY};font-size:12px;'>Opponents: 3 remaining, avg budget ~{sim_budget}cr</span>",
+                    unsafe_allow_html=True)
+
+    if st.button("🎯 Run Live Auction 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:
+                idx = player_pool.head(60)[player_pool.head(60)["name"] == sim_player].index
+                if len(idx) > 0:
+                    i = list(player_pool.head(60).index).index(idx[0])
+                    p10 = qe_result["preds"]["P10"][i]
+                    p50 = qe_result["preds"]["P50"][i]
+                    p90 = qe_result["preds"]["P90"][i]
+                else:
+                    p10, p50, p90 = target["projected_points"] * 0.8, target["projected_points"], target["projected_points"] * 1.2
+            else:
+                p10, p50, p90 = target["projected_points"] * 0.8, target["projected_points"], target["projected_points"] * 1.2
+
+            pv = PlayerValuation(
+                name=target["name"], team=target["team"], role=target["role"],
+                projected_points=target["projected_points"],
+                market_value=target["market_value"],
+                ceiling_price=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": sim_round,
+                "total_rounds": 20,
+                "opponent_budgets": [sim_budget, sim_budget - 50, sim_budget + 30],
+                "players_remaining_in_role": {"P": 5, "D": 10, "C": 10, "A": 8},
+                "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 = {
+                "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).values[0])
+            win_pct = bandit_bid / max(bandit_bid + opp_bid, 1) * 100
+
+            bo = BudgetOptimizer(total_budget=500)
+            role_budget = bo.optimize(player_pool.head(30), n_calls=5)
+            role_rec = role_budget.get(target["role"], 50)
+
+            with st.container():
+                st.markdown("### 📊 Decision Analysis")
+
+                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("RECOMMENDED BID", f"{bandit_bid} cr",
+                                          f"vs opponent ~{opp_bid:.0f} cr", bid_color),
+                                unsafe_allow_html=True)
+                with kd2:
+                    st.markdown(kpi_card("WIN PROBABILITY", f"{min(win_pct, 95):.0f}%",
+                                          "", GOLD), unsafe_allow_html=True)
+                with kd3:
+                    st.markdown(kpi_card("P50 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"for {target['role']} players", VIOLET),
+                                unsafe_allow_html=True)
+                with kd5:
+                    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)
+                    st.markdown(kpi_card("DOWNSIDE RISK", risk_label,
+                                          f"VaR floor: {p10:.1f}", risk_color),
+                                unsafe_allow_html=True)
+
+                info_msg = (
+                    f"**{target['name']}** ({ROLE_ICONS.get(target['role'], '')} {target['role']}) — "
+                    f"Bandit recommends **{bandit_bid} cr** bid. Opponents likely to bid ~{opp_bid:.0f} cr. "
+                    f"Budget optimizer allocates ~{role_rec:.0f} cr for {target['role']} role."
+                )
+                if bandit_bid > opp_bid:
+                    info_msg += f"\n\n🟢 Expected to win this player. Value ratio: {target['projected_points'] / max(bandit_bid, 1):.3f} FV/cr."
+                else:
+                    info_msg += f"\n\n🔴 Opponent likely to outbid. Consider increasing bid or skipping for better value."
+                insight(info_msg)
+
+        except Exception as e:
+            st.error(f"Simulation error: {e}")
+
+    st.divider()
+
+    # ── Phase-by-phase showcase ──
+    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("⚡ Phase 1 — Prediction Quality: Quantile Ensemble")
+        if "error" in results.get("quantile", {}):
+            st.warning(f"Quantile model error: {results['quantile']['error']}")
+        else:
+            q = results["quantile"]
+            fig = _phase_chart_quantile(q["preds"], q["risk"])
+            st.plotly_chart(fig, width="stretch")
+            insight("P10/P50/P90 predictions enable VaR-constrained bidding. "
+                     "The risk histogram shows P(FV < 5.5) per player — yellow cards kill your matchday score.")
+
+        section("⏱ Phase 1 — Prediction Quality: Minutes Survival Model")
+        if "error" in results.get("survival", {}):
+            st.warning(f"Survival model error: {results['survival']['error']}")
+        else:
+            s = results["survival"]
+            fig = _phase_chart_survival(s["expected"], s["lower"], s["upper"], s["starter_probs"])
+            st.plotly_chart(fig, width="stretch")
+            insight("Weibull AFT predicts full minutes distribution — not just binary starter flag. "
+                     "A player projected at 7.5 who only plays 60% of matches is auction poison.")
+
+    with tab2:
+        section("🎰 Phase 2 — Thompson Sampling Bandit for Live Bidding")
+        if "error" in results.get("bandit", {}):
+            st.warning(f"Bandit model error: {results['bandit']['error']}")
+        else:
+            fig = _phase_chart_bandit(results["bandit"]["trial_bids"])
+            st.plotly_chart(fig, width="stretch")
+            insight("Thompson Sampling learns optimal bid levels per role over time. "
+                     "Explores cheap sleepers when uncertainty is high, exploits known stars when confident.")
+
+        section("💰 Phase 2 — Opponent Bidding Model")
+        if "error" in results.get("opponent_bidding", {}):
+            st.warning(f"Opponent bidding error: {results['opponent_bidding']['error']}")
+        else:
+            ob = results["opponent_bidding"]
+            fig = _phase_chart_opponent_bids(ob["sample_bids"])
+            st.plotly_chart(fig, width="stretch")
+            insight("LightGBM predicts opponent max bids per player. Outbid intelligently — "
+                     "don't overpay when no competitor is interested.")
+
+        section("📐 Phase 2 — Bayesian Budget Optimization")
+        if "error" in results.get("budget_opt", {}):
+            st.warning(f"Budget optimizer error: {results['budget_opt']['error']}")
+        else:
+            fig = _phase_chart_budget_opt(results["budget_opt"]["allocation"])
+            st.plotly_chart(fig, width="stretch")
+            insight("Gaussian Process optimization finds the optimal budget split across roles. "
+                     "GK gets ~8%, DEF ~35%, MID ~32%, FWD ~25% — adapts to pool quality.")
+
+    with tab3:
+        section("🧠 Phase 3 — Double DQN Auction Agent")
+        if "error" in results.get("rl", {}):
+            st.warning(f"RL agent error: {results['rl']['error']}")
+        else:
+            rl = results["rl"]
+            k_rl1, k_rl2, k_rl3 = st.columns(3)
+            with k_rl1:
+                st.markdown(kpi_card("STATE DIM", str(rl["observation_dim"]), "features", SKY), unsafe_allow_html=True)
+            with k_rl2:
+                st.markdown(kpi_card("ACTION SPACE", str(rl["action_dim"]), "bid levels", GOLD), unsafe_allow_html=True)
+            with k_rl3:
+                st.markdown(kpi_card("EXPERIENCES", str(rl["buffer_size"]), "stored", VIOLET), unsafe_allow_html=True)
+            insight("Double DQN agent trained on 20 episodes (fast demo). In production, train 5000+ episodes "
+                     "against diverse simulated opponents. Learns to delay big bids until rivals are exhausted.")
+
+        section("🧩 Phase 3 — Set Transformer: Team Value ≠ Sum of Parts")
+        if "error" in results.get("set", {}):
+            st.warning(f"Set Transformer error: {results['set']['error']}")
+        else:
+            sf = results["set"]
+            s_k1, s_k2, s_k3 = st.columns(3)
+            with s_k1:
+                st.markdown(kpi_card("BASE TEAM VALUE", f"{sf['base_value']:.0f} PTS",
+                                      "sum of projections", SKY), unsafe_allow_html=True)
+            with s_k2:
+                delta_color = PITCH_GREEN if sf["marginal_value"] > 0 else RED
+                st.markdown(kpi_card("MARGINAL PLAYER", f"{sf['marginal_value']:+.1f} PTS",
+                                      "value added by 1 player", delta_color), unsafe_allow_html=True)
+            with s_k3:
+                red_color = PITCH_GREEN if sf["redundancy"] < 0.5 else GOLD
+                st.markdown(kpi_card("REDUNDANCY", f"{sf['redundancy']:.2f}",
+                                      "0=diverse 1=overlapping", red_color), unsafe_allow_html=True)
+            insight("Set Transformer captures non-linear synergies. Two playmakers overlapping = "
+                     "worth less than sum of parts. Redundancy score warns you before overpaying.")
+
+    with tab4:
+        section("📊 Phase 4 — Hierarchical Bayesian Pooling")
+        if "error" in results.get("bayesian", {}):
+            st.warning(f"Bayesian model error: {results['bayesian']['error']}")
+        else:
+            b = results["bayesian"]
+            fig = _phase_chart_bayesian(b["mean"], b["std"], b["reliability"])
+            st.plotly_chart(fig, width="stretch")
+            insight("Hierarchical model shrinks rookies toward role mean. Low reliability = high uncertainty. "
+                     "Don't pay premium prices for players with <10 career games.")
+
+        section("🎯 Phase 4 — Conformal Prediction Bands")
+        if "error" in results.get("conformal", {}):
+            st.warning(f"Conformal predictor error: {results['conformal']['error']}")
+        else:
+            cp_r = results["conformal"]
+            fig = _phase_chart_conformal(cp_r["predictions"], cp_r["lowers"], cp_r["uppers"], cp_r["coverage"])
+            st.plotly_chart(fig, width="stretch")
+            insight(f"Conformal bands: {cp_r['coverage']*100:.0f}% coverage on test set. "
+                     "Model-agnostic calibrated intervals — no distributional assumptions needed.")
+
+    with tab5:
+        section("🔗 Phase 5 — Graph Attention Network: Player Chemistry")
+        if "error" in results.get("chemistry", {}):
+            st.warning(f"Chemistry model error: {results['chemistry']['error']}")
+        else:
+            fig = _phase_chart_chemistry(results["chemistry"]["bonus_matrix"])
+            st.plotly_chart(fig, width="stretch")
+            insight("Pairwise chemistry bonuses from pass networks, assists, and crosses. "
+                     "A strong winger→striker edge boosts both players. Target linked pairs in the auction.")
+
+        section("🔥 Phase 5 — Hawkes Process: Form Momentum")
+        if "error" in results.get("hawkes", {}):
+            st.warning(f"Form model error: {results['hawkes']['error']}")
+        else:
+            fig = _phase_chart_hawkes(results["hawkes"]["statuses"])
+            st.plotly_chart(fig, width="stretch")
+            insight("Self-exciting process detects HOT/COLD streaks. A HOT player on a 5-game scoring run "
+                     "has temporarily elevated projection. Exploit recency bias in your opponents.")
+
+    with tab6:
+        section("🔬 Phase 6 — Causal Forest: Transfer Effect Analysis")
+        if "error" in results.get("causal", {}):
+            st.warning(f"Causal forest error: {results['causal']['error']}")
+        else:
+            cf = results["causal"]
+            k_c1, k_c2, k_c3 = st.columns(3)
+            with k_c1:
+                color = PITCH_GREEN if cf["ate"] > 0 else RED
+                st.markdown(kpi_card("AVG TREATMENT EFFECT", f"{cf['ate']:+.3f}",
+                                      "adding a player", color), unsafe_allow_html=True)
+            with k_c2:
+                st.markdown(kpi_card("CI LOWER", f"{cf['ate_lower']:+.3f}", "95% confidence", TEXT_SECONDARY),
+                            unsafe_allow_html=True)
+            with k_c3:
+                st.markdown(kpi_card("CI UPPER", f"{cf['ate_upper']:+.3f}", "95% confidence", TEXT_SECONDARY),
+                            unsafe_allow_html=True)
+            insight("Causal forest estimates the TRUE effect of a roster change, controlling for confounders. "
+                     "Adding a top midfielder doesn't help if you already have 5 strong mids — "
+                     "the diminishing returns are captured in the CATE.")
+
+    # ── Methodology ──
+    st.divider()
+    with st.expander("⚙️ Methodology — 10 ML Models Explained"):
+        st.markdown("""
+        | # | Model | Type | What It Does |
+        |---|-------|------|-------------|
+        | 1 | QuantileEnsemble | LightGBM quantile | P10/P50/P90 predictions → risk-aware bidding |
+        | 2 | MinutesSurvivalModel | Weibull AFT | Full minutes distribution, starter probability |
+        | 3 | BanditAuctionSolver | Thompson Sampling | Optimal bid per round balancing explore/exploit |
+        | 4 | OpponentBidModel | LightGBM regressor | Predicts competitor max bid per player |
+        | 5 | BudgetOptimizer | Bayesian Optimization | Optimal budget split across GK/DEF/MID/FWD |
+        | 6 | PlayerChemistryGAT | Graph Attention Network | Player synergy bonuses from pass networks |
+        | 7 | PlayerFormModel | Hawkes Process | Momentum/decorrelation hot streak detection |
+        | 8 | BayesianPlayerModel | Hierarchical Bayes | Rookie uncertainty via role-level shrinkage |
+        | 9 | ConformalPredictor | Conformal inference | Calibrated prediction bands, model-agnostic |
+        | 10 | SetTransformer | Transformer on sets | Team composition value beyond sum-of-parts |
+        | 11 | RLAuctionPolicy | Double DQN | RL agent for sequential auction strategy |
+        | 12 | TransferCausalModel | Causal Forest | Causal effect of transfer on team performance |
+        """, unsafe_allow_html=False)
+
+    st.divider()
+    st.caption("Dev Preview v1.0 — all models running on synthetic data. Connect real pipeline for production use.")
+
+
+if __name__ == "__main__":
+    run()

dashboard/warehouse.py

@@ -2,40 +2,59 @@
 Cached with @st.cache_data. No imports from ML code.
 """
 
+import logging
 from pathlib import Path
 
 import pandas as pd
 
+logger = logging.getLogger(__name__)
+
 ROOT = Path(__file__).resolve().parent.parent
 WAREHOUSE = ROOT / "data" / "warehouse"
 
+_EMPTY_DEFAULTS = {
+    "players.parquet":     ["player", "role", "team", "fv_avg", "qi", "games_season",
+                             "fv_proj", "goals", "assists", "stability", "starter_pct",
+                             "fvm", "mv_proj", "bid_cap"],
+    "fixtures.parquet":    ["team", "matchday", "opp_strength", "home", "away"],
+    "predictions.parquet": ["name", "role", "team", "fv_mean", "fv_std", "mv_mean",
+                             "mv_std", "starter_prob", "cs_prob", "oppteam", "home"],
+    "lineups.parquet":     ["player", "role", "team", "starter_pct"],
+    "votes.parquet":       ["player", "vote", "matchday"],
+    "model_metrics.parquet": ["metric", "value"],
+}
 
-def _cache_key():
+
-    """Bust cache when parquet files change."""
+def _read_parquet_or_empty(name):
-    files = sorted(WAREHOUSE.glob("*.parquet"))
+    path = WAREHOUSE / name
-    mtimes = tuple(f.stat().st_mtime for f in files)
+    if path.exists():
-    return (len(files), mtimes)
+        return pd.read_parquet(path)
+    cols = _EMPTY_DEFAULTS.get(name, [])
+    df = pd.DataFrame([{c: (0.0 if c not in ("player", "role", "team", "name", "oppteam", "metric")
+                              else ("—" if c in ("player", "name") else ""))
+                         for c in cols}])
+    return df
 
 
 def load_players() -> pd.DataFrame:
-    return pd.read_parquet(WAREHOUSE / "players.parquet")
+    return _read_parquet_or_empty("players.parquet")
 
 
 def load_fixtures() -> pd.DataFrame:
-    return pd.read_parquet(WAREHOUSE / "fixtures.parquet")
+    return _read_parquet_or_empty("fixtures.parquet")
 
 
 def load_predictions() -> pd.DataFrame:
-    return pd.read_parquet(WAREHOUSE / "predictions.parquet")
+    return _read_parquet_or_empty("predictions.parquet")
 
 
 def load_lineups() -> pd.DataFrame:
-    return pd.read_parquet(WAREHOUSE / "lineups.parquet")
+    return _read_parquet_or_empty("lineups.parquet")
 
 
 def load_votes() -> pd.DataFrame:
-    return pd.read_parquet(WAREHOUSE / "votes.parquet")
+    return _read_parquet_or_empty("votes.parquet")
 
 
 def load_model_metrics() -> pd.DataFrame:
-    return pd.read_parquet(WAREHOUSE / "model_metrics.parquet")
+    return _read_parquet_or_empty("model_metrics.parquet")