""" Corrected re-run of research.ipynb (binned_df path). Faithfully reproduces the notebook pipeline (cells 3-24: load live_data, highest-CLV pick, time-bin target selection from first 20%, exclude USL/BSA, dedup one snapshot per game near TARGET_MINS) then applies ONLY the EVALUATION fixes (core strategy untouched: draws KEPT, original Kalshi-CLV floor >0.04, NO upper cap): 1. NET returns: win = decimal_odds - 1 (not the gross decimal odds). 2. Significance run on the NET returns (correct breakeven = 0). 3. Kalshi Kelly uses pin_fair of the bet side as p_win (not the market ask). """ import os, math, sys import pandas as pd, numpy as np from scipy import stats import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt sys.stdout.reconfigure(encoding='utf-8', errors='replace') os.chdir(r'E:\soccer') # ---- load (cell 3-4) ---- dfs = [] for f in os.listdir('live_data'): if f.endswith('.csv') and f != 'test.csv': dfs.append(pd.read_csv('live_data/' + f)) df = pd.concat(dfs) # ---- highest-CLV pick + correctness (cells 5-6) ---- def highest_clv(h, d, a): cd = {'H': h, 'D': d, 'A': a} if pd.isna(cd['H']): return None, np.nan k = max(cd, key=cd.get) return k, cd[k] for bk in ('b365', 'kalshi'): df[f'{bk}_highest_clv'] = df.apply( lambda r: highest_clv(r[f'clv_{bk}_home'], r[f'clv_{bk}_draw'], r[f'clv_{bk}_away'])[0], axis=1) df[f'{bk}_highest_clv_val'] = df.apply( lambda r: highest_clv(r[f'clv_{bk}_home'], r[f'clv_{bk}_draw'], r[f'clv_{bk}_away'])[1], axis=1) df['correct_b365'] = (df['ftr'] == df['b365_highest_clv']).astype(int) df['correct_kalshi'] = (df['ftr'] == df['kalshi_highest_clv']).astype(int) # ---- time-bin target selection from first 20% (cells 18-24) ---- dd = df.dropna(subset=['b365_highest_clv', 'kalshi_highest_clv']) dd = dd[dd['status'] == 'PRE'] num_bins = math.ceil(abs(int(dd['elapsed'].min())) / 240) bins = np.histogram(dd['elapsed'], bins=num_bins)[1] def assign_bin(x): for i in range(len(bins) - 1): if bins[i] <= x <= bins[i + 1]: return i dd['bin'] = dd['elapsed'].apply(assign_bin) first_20 = round(dd.shape[0] * 0.2) df_first = dd.head(first_20) binned = (df_first[['correct_kalshi', 'bin']].groupby('bin') .agg({'correct_kalshi': ['mean', 'count']}).reset_index()) binned['bin_mid'] = binned['bin'].apply(lambda x: round((bins[x] + bins[x + 1]) / 2)) best = binned.sort_values(('correct_kalshi', 'mean'), ascending=False).iloc[0] TARGET_MINS = int(best['bin_mid'].iloc[0]) if hasattr(best['bin_mid'], 'iloc') else int(best['bin_mid']) df_rem = dd.iloc[first_20:].copy() df_rem = df_rem[(df_rem['league'] != 'USL') & (df_rem['league'] != 'BSA')] df_rem['dist'] = (df_rem['elapsed'] - TARGET_MINS).abs() df_rem = df_rem.sort_values('dist').drop_duplicates(['home', 'away'], keep='first') df_rem['timestamp'] = df_rem['timestamp'].apply(lambda x: pd.to_datetime(x).timestamp()) g = df_rem.sort_values('timestamp').reset_index(drop=True) g = g.dropna(subset=['result_home', 'result_away', 'clv_kalshi_home', 'clv_kalshi_draw', 'clv_kalshi_away']) print(f"TARGET_MINS (best bin from first 20%): {TARGET_MINS}") print(f"games after dedup/exclude USL+BSA: {len(g)}\n") # ---- Kelly engine (cells 12-13) ---- side = {'H': '_home', 'D': '_draw', 'A': '_away'} def kelly_size(p, odds, br, mf=0.025): b = odds - 1 f = (p * odds - (1 - p)) / b return max(0.0, min(f, mf)) * br def compound(odds, probs, outcomes, br0=1000.0, mf=0.025): br = br0; eq = [] for o, p, w in zip(odds, probs, outcomes): bet = kelly_size(p, o, br, mf) br += bet * (o - 1) if w else -bet eq.append(br) return eq FLOOR = 0.04 # original notebook floor; draws KEPT, no upper cap (core strategy unchanged) gate = g[g['kalshi_highest_clv_val'] > FLOOR] def evaluate(name, pick_col, clv_val_col, correct_col, odds_fn, pinfair_fn): """Strategy on its own pick side; original Kalshi-CLV gate (>0.04), draws kept.""" s = gate.copy() odds = s.apply(odds_fn, axis=1) pin = s.apply(pinfair_fn, axis=1) won = s[correct_col].values net = np.where(won == 1, odds - 1, -1.0) # FIX: NET returns eq = compound(odds.values, pin.values, won) # FIX: pin_fair p_win # FIX: significance on NET returns z = net.mean() / (net.std(ddof=1) / np.sqrt(len(net))) p = 1 - stats.norm.cdf(z) wr = won.mean() return dict(name=name, n=len(s), wr=wr, net_sum=net.sum(), roi=100 * net.mean(), z=z, p=p, final=eq[-1], eq=eq) results = [ evaluate('B365', 'b365_highest_clv', 'b365_highest_clv_val', 'correct_b365', lambda r: r['b365' + side[r['b365_highest_clv']]], lambda r: r['pin_fair' + side[r['b365_highest_clv']]]), evaluate('Kalshi', 'kalshi_highest_clv', 'kalshi_highest_clv_val', 'correct_kalshi', lambda r: 1.0 / r['kalshi' + side[r['kalshi_highest_clv']] + '_ask'], lambda r: r['pin_fair' + side[r['kalshi_highest_clv']]]), evaluate('Kalshi w/ B365', 'b365_highest_clv', 'b365_highest_clv_val', 'correct_b365', lambda r: 1.0 / r['kalshi' + side[r['b365_highest_clv']] + '_ask'], lambda r: r['pin_fair' + side[r['b365_highest_clv']]]), ] print("=== CORRECTED (net returns, pin_fair Kelly) | core unchanged: draws KEPT, floor>0.04, no cap ===") print(f"{'strategy':>16}{'N':>5}{'win%':>7}{'netSum':>9}{'ROI%':>8}{'z':>7}{'p':>8}{'Kelly$':>10}") print('-' * 71) for r in results: sig = '*' if r['p'] < 0.05 else ' ' print(f"{r['name']:>16}{r['n']:>5}{100*r['wr']:>6.1f}%{r['net_sum']:>+9.1f}" f"{r['roi']:>+8.1f}{r['z']:>+7.2f}{r['p']:>8.4f}{sig}{r['final']:>9.0f}") print("\n(Kelly$ = final compounded bankroll from $1000, real Kelly w/ pin_fair, 2.5% cap)") # ---- equity chart ---- colors = {'B365': '#2077b4', 'Kalshi': '#d62728', 'Kalshi w/ B365': '#2ca02c'} fig, ax = plt.subplots(figsize=(11, 5.5)) for r in results: ax.plot(range(len(r['eq'])), r['eq'], color=colors[r['name']], lw=2, label=f"{r['name']} (${r['final']:,.0f}, n={r['n']})") ax.axhline(1000, color='gray', ls='--', lw=1, alpha=0.6) ax.set_xlabel('Bets'); ax.set_ylabel('Bankroll ($)') ax.set_title('Corrected compounded returns (net, pin_fair Kelly) - draws kept, no cap') ax.legend(loc='best'); ax.grid(True, alpha=0.3); fig.tight_layout() fig.savefig('test_fixed.png', dpi=120); plt.close(fig) print("\nchart: test_fixed.png")