"""MOD-EJEC · 30 · implementación 1 del PREREG_1 (md5 ddf7da57…): P(modificación inscrita en el RPC en los 12 meses siguientes a la adjudicación). Pruebas t = 2024 y t = 2023, entrenamiento 2018..t−1. -> salidas/30_rpc.json, priv/pred_rpc_.csv""" import json, time, sys sys.path.insert(0, '/w/py') import numpy as np, pandas as pd, sklearn from scipy.special import logit from sklearn.linear_model import LogisticRegression from sklearn.model_selection import GroupKFold import statsmodels.api as sm from rpc_comun import leer, historial_mes, variables, diseno, cadena, NUM, CATS, C_L2 T0 = time.time(); SEM = 20261104; B = 2000 d = leer(); H = historial_mes(d); F = variables(d, H) H.to_csv('/w/priv/hist_rpc.csv'); print('historial', round(time.time() - T0), 's', flush=True) y = d['y'].to_numpy(float); grp = d['org'].fillna('∅').to_numpy() def auc(s, yy, ww=None): ww = np.ones(len(s)) if ww is None else ww u, g = np.unique(s, return_inverse=True); n = len(u) wp = np.bincount(g, ww * yy, n); wn = np.bincount(g, ww * (1 - yy), n); below = np.cumsum(wn) - wn return float(np.sum(wp * (below + 0.5 * wn)) / (wp.sum() * wn.sum())) def ajustar(X, yy): return LogisticRegression(C=C_L2, solver='lbfgs', max_iter=5000).fit(X, yy) R = {'prerregistro': 'ddf7da57a51b91953cf929195ac62a15', 'versiones': {'sklearn': sklearn.__version__, 'numpy': np.__version__, 'pandas': pd.__version__}, 'variables': NUM, 'categoricas': CATS, 'C': C_L2} for t in (2024, 2023): TR = ((d['T'] >= 2018) & (d['T'] <= t - 1)).to_numpy(); TE = (d['T'] == t).to_numpy() X, info = diseno(d, F, TR); Xtr, Xte = X(TR), X(TE) m = ajustar(Xtr, y[TR]) oof = np.zeros(TR.sum()) for a, b in GroupKFold(n_splits=5).split(Xtr, y[TR], grp[TR]): oof[b] = ajustar(Xtr[a], y[TR][a]).predict_proba(Xtr[b])[:, 1] platt = LogisticRegression(C=1e6).fit(logit(np.clip(oof, 1e-6, 1 - 1e-6))[:, None], y[TR]) p_raw = m.predict_proba(Xte)[:, 1]; p = platt.predict_proba(logit(np.clip(p_raw, 1e-6, 1 - 1e-6))[:, None])[:, 1] b1 = cadena(d.loc[TR], y[TR], d.loc[TE], (('amb', ['ambit']), ('o', ['ambit', 'org']), ('oc', ['ambit', 'org', 'cpv_div'])))['oc'][0] li = lambda mask: np.log10(np.maximum(d.loc[mask, 'imp_adj'], 1)).to_numpy()[:, None] b3 = LogisticRegression(C=1e6).fit(li(TR), y[TR]).predict_proba(li(TE))[:, 1] yt = y[TE]; cl = pd.factorize(d.loc[TE, 'org'].fillna('∅'))[0]; Cn = cl.max() + 1 pt = {'p': auc(p, yt), 'p_raw': auc(p_raw, yt), 'b1': auc(b1, yt), 'b3': auc(b3, yt)} rep = {k: [] for k in pt}; rng = np.random.default_rng(SEM) for _ in range(B): mult = np.bincount(rng.integers(0, Cn, Cn), minlength=Cn)[cl].astype(float) if (yt * mult).sum() == 0 or ((1 - yt) * mult).sum() == 0: continue for k, s in (('p', p), ('p_raw', p_raw), ('b1', b1), ('b3', b3)): rep[k].append(auc(s, yt, mult)) rep = {k: np.array(v) for k, v in rep.items()} ic = {k: [float(np.percentile(v, 2.5)), float(np.percentile(v, 97.5))] for k, v in rep.items()} dl = {k: {'delta': pt['p'] - pt[k], 'ic': [float(np.percentile(rep['p'] - rep[k], 2.5)), float(np.percentile(rep['p'] - rep[k], 97.5))]} for k in ('b1', 'b3')} sl = sm.GLM(yt, sm.add_constant(logit(np.clip(p, 1e-6, 1 - 1e-6))), family=sm.families.Binomial()).fit() q = pd.qcut(p, 10, labels=False, duplicates='drop') dec = pd.DataFrame({'p': p, 'y': yt, 'q': q}).groupby('q').agg(n=('p', 'size'), predicha=('p', 'mean'), observada=('y', 'mean')).reset_index() dec['dif_puntos'] = 100 * (dec['predicha'] - dec['observada']); nd = len(dec) cal_ok = bool((dec['dif_puntos'].abs().iloc[:nd - 1] <= 5).all() and abs(dec['dif_puntos'].iloc[-1]) <= 8) orden = np.argsort(-p); top = {f'top{k}': float(yt[orden[:int(len(p) * k / 100)]].mean()) for k in (5, 10)} yimp = d.loc[TE, 'y_imp'].to_numpy(float) amb = d.loc[TE, 'ambit'].fillna('∅').to_numpy() por_ambit = {a: {'n': int((amb == a).sum()), 'pos': int(yt[amb == a].sum()), 'auc': auc(p[amb == a], yt[amb == a]) if 0 < yt[amb == a].sum() < (amb == a).sum() else None} for a in sorted(set(amb))} R[str(t)] = {'n_train': int(TR.sum()), 'pos_train': int(y[TR].sum()), 'n_test': int(TE.sum()), 'pos_test': int(yt.sum()), 'tasa_test': float(yt.mean()), 'auc': pt, 'auc_ic': ic, 'delta': dl, 'o_e': float(yt.sum() / p.sum()), 'pendiente': float(sl.params[1]), 'brier': float(np.mean((p - yt) ** 2)), 'deciles': dec.to_dict(orient='records'), 'calibracion_ok': cal_ok, 'precision_superior': top, 's3_auc_con_efecto_economico': auc(p, yimp), 's4_por_ambito': por_ambit, 'platt': {'a': float(platt.coef_[0][0]), 'b': float(platt.intercept_[0])}, 'niveles': {c: len(v) for c, v in info['niv'].items()}} out = d.loc[TE, ['id_canonical', 'org', 'ambit', 'T', 'y', 'y_imp']].copy(); out['p'] = p; out['p_raw'] = p_raw; out['b1'] = b1; out['b3'] = b3 out.to_csv(f'/w/priv/pred_rpc_{t}.csv', index=False) r = R[str(t)] print(t, 'AUC', round(pt['p'], 4), ic['p'], 'B1', round(pt['b1'], 4), 'Δ', round(dl['b1']['delta'], 4), 'B3', round(pt['b3'], 4), 'O/E', round(r['o_e'], 3), 'pend', round(r['pendiente'], 3), 'cal_ok', cal_ok, 'max dif', round(float(dec['dif_puntos'].abs().max()), 2), top, round(time.time() - T0), 's', flush=True) a24, a23 = R['2024']['auc']['p'], R['2023']['auc']['p'] crit = {'(1) AUC 2024 >= 0,68 y 2023 >= 0,66': a24 >= 0.68 and a23 >= 0.66, '(2) calibracion en los dos años': R['2024']['calibracion_ok'] and R['2023']['calibracion_ok'], '(3) delta B1 >= 0,015 en los dos años': R['2024']['delta']['b1']['delta'] >= 0.015 and R['2023']['delta']['b1']['delta'] >= 0.015, 'publicacion: AUC >= 0,75 en los dos años': a24 >= 0.75 and a23 >= 0.75} R['criterios_1_a_3'] = crit print(json.dumps(crit, ensure_ascii=False)) json.dump(R, open('/w/salidas/30_rpc.json', 'w'), ensure_ascii=False, indent=1, default=float)