#!/usr/bin/env python3 """ 2026年1-4月每日新增注册人数回归分析（V4 - 比例潮汐+保底） 改进点： 1. 星期潮汐效应改为比例补偿：周末效应 = 趋势 × (1 + 比例)，而非趋势 + 固定值 理由：1月基线50人时+11是+22%，4月基线120人时+11只+9%，固定值会随趋势增长等比稀释 2. 实际值低于拟合值时保留实际值（活动日才用拟合值替换） """ import numpy as np import pandas as pd # ==================== 全量数据 ==================== raw_data = { '2026-01-01': 78, '2026-01-02': 74, '2026-01-03': 69, '2026-01-04': 44, '2026-01-05': 56, '2026-01-06': 33, '2026-01-07': 52, '2026-01-08': 60, '2026-01-09': 58, '2026-01-10': 84, '2026-01-11': 75, '2026-01-12': 35, '2026-01-13': 46, '2026-01-14': 59, '2026-01-15': 32, '2026-01-16': 37, '2026-01-17': 67, '2026-01-18': 71, '2026-01-19': 53, '2026-01-20': 49, '2026-01-21': 42, '2026-01-22': 63, '2026-01-23': 48, '2026-01-24': 72, '2026-01-25': 86, '2026-01-26': 62, '2026-01-27': 58, '2026-01-28': 150, '2026-01-29': 106, '2026-01-30': 88, '2026-01-31': 61, '2026-02-01': 55, '2026-02-02': 53, '2026-02-03': 42, '2026-02-04': 39, '2026-02-05': 42, '2026-02-06': 55, '2026-02-07': 37, '2026-02-08': 47, '2026-02-09': 42, '2026-02-10': 60, '2026-02-11': 265, '2026-02-12': 59, '2026-02-13': 42, '2026-02-14': 32, '2026-02-15': 41, '2026-02-16': 28, '2026-02-17': 52, '2026-02-18': 23, '2026-02-19': 35, '2026-02-20': 26, '2026-02-21': 37, '2026-02-22': 36, '2026-02-23': 47, '2026-02-24': 45, '2026-02-25': 64, '2026-02-26': 182, '2026-02-27': 102, '2026-02-28': 168, '2026-03-01': 126, '2026-03-02': 108, '2026-03-03': 78, '2026-03-04': 111, '2026-03-05': 453, '2026-03-06': 218, '2026-03-07': 241, '2026-03-08': 207, '2026-03-09': 168, '2026-03-10': 109, '2026-03-11': 135, '2026-03-12': 178, '2026-03-13': 245, '2026-03-14': 131, '2026-03-15': 132, '2026-03-16': 110, '2026-03-17': 96, '2026-03-18': 96, '2026-03-19': 78, '2026-03-20': 103, '2026-03-21': 133, '2026-03-22': 126, '2026-03-23': 85, '2026-03-24': 80, '2026-03-25': 94, '2026-03-26': 85, '2026-03-27': 95, '2026-03-28': 105, '2026-03-29': 101, '2026-03-30': 82, '2026-03-31': 100, '2026-04-01': 138, '2026-04-02': 165, '2026-04-03': 779, '2026-04-04': 341, '2026-04-05': 186, '2026-04-06': 207, '2026-04-07': 212, '2026-04-08': 751, '2026-04-09': 341, '2026-04-10': 129, '2026-04-11': 174, '2026-04-12': 151, '2026-04-13': 118, '2026-04-14': 133, '2026-04-15': 126, '2026-04-16': 115, '2026-04-17': 84, '2026-04-18': 117, '2026-04-19': 120, '2026-04-20': 89, '2026-04-21': 100, '2026-04-22': 179, '2026-04-23': 139, '2026-04-24': 122, '2026-04-25': 142, '2026-04-26': 137, '2026-04-27': 120, '2026-04-28': 164, '2026-04-29': 65, '2026-04-30': 58, } dates = sorted(raw_data.keys()) date_series = pd.to_datetime(dates) values = np.array([raw_data[d] for d in dates], dtype=float) n = len(values) x = np.arange(n, dtype=float) dow = np.array([d.weekday() for d in date_series]) # 0=Mon, 6=Sun # ==================== 活动标记 ==================== activity_events = { '2026-01-28': ('活动', 1), '2026-02-11': ('活动', 0), '2026-02-26': ('活动', 4), '2026-03-05': ('活动', 3), '2026-03-09': ('活动', 0), '2026-03-12': ('活动', 0), '2026-03-13': ('活动', 0), '2026-04-03': ('大型活动', 4), '2026-04-08': ('大型活动', 2), '2026-04-22': ('小活动', 1), '2026-04-28': ('小活动', 0), } affected_reason = {} for d_str, (reason, aftereffect_days) in activity_events.items(): base_idx = dates.index(d_str) affected_reason[d_str] = reason for offset in range(1, aftereffect_days + 1): idx = base_idx + offset if idx < n: ae_date = dates[idx] if ae_date not in affected_reason: affected_reason[ae_date] = f'{dates[base_idx]} 余波' is_affected = np.array([d in affected_reason for d in dates]) # ==================== LOESS ==================== def loess_smooth(x_data, y_data, x_eval, frac=0.28, deg=2): n_eval = len(x_eval) n_data = len(x_data) y_smooth = np.zeros(n_eval) for j in range(n_eval): x0 = x_eval[j]; dist = np.abs(x_data - x0) k = max(int(frac * n_data), deg + 2); k = min(k, n_data) idx = np.argpartition(dist, k)[:k] x_nn, y_nn, d_nn = x_data[idx], y_data[idx], dist[idx] max_dist = np.max(d_nn) w = (1 - (d_nn / max_dist) ** 3) ** 3 if max_dist > 0 else np.ones(k) A = np.column_stack([np.ones(k), x_nn, x_nn**2]); W = np.diag(w) try: beta = np.linalg.solve(A.T @ W @ A, A.T @ W @ y_nn) y_smooth[j] = beta[0] + beta[1] * x0 + beta[2] * x0**2 except np.linalg.LinAlgError: y_smooth[j] = np.average(y_nn, weights=w) return np.maximum(y_smooth, 0) clean_mask = ~is_affected clean_x, clean_y, clean_dow = x[clean_mask], values[clean_mask], dow[clean_mask] # 第一轮：LOESS 拟合原始趋势 trend_raw = loess_smooth(clean_x, clean_y, x, frac=0.28, deg=2) # ==================== 星期效应：比例方式 ==================== # 在干净数据上计算残差率 = (实际 - 趋势) / 趋势 residual_pct = (clean_y - trend_raw[clean_mask]) / trend_raw[clean_mask] dow_ratio = np.zeros(7) for d in range(7): mask = clean_dow == d if np.sum(mask) > 0: dow_ratio[d] = np.mean(residual_pct[mask]) # 调整使比例总和为0（不影响总注册人数） dow_ratio -= np.mean(dow_ratio) print("=" * 80) print("2026年3-4月 每日新增注册回归分析 V4（比例潮汐 + 保底）") print("=" * 80) print(f"\n📊 星期潮汐效应（比例方式）：") dow_names = ['周一', '周二', '周三', '周四', '周五', '周六', '周日'] for i, name in enumerate(dow_names): bar = '█' * max(1, int(abs(dow_ratio[i]) * 200)) sign = '+' if dow_ratio[i] >= 0 else '' print(f" {name}: {sign}{dow_ratio[i]*100:.1f}% {bar}") # ==================== 最终拟合 = 趋势 × (1 + 星期比例) ==================== trend_with_dow = trend_raw * (1 + dow_ratio[dow]) trend_with_dow = np.maximum(trend_with_dow, 0) # ==================== 有效注册人数 ==================== # 规则：活动日 → min(实际, 拟合)；非活动日 → 保留实际值 effective = np.where(is_affected, np.minimum(values, trend_with_dow), values) print(f"\n{'='*80}") print("📋 2026年3-4月 最终结果") print(f"{'='*80}") print(f"{'日期':>12} {'星期':>4} | {'实际':>5} | {'影响':>10} | {'趋势':>6} | {'趋势+潮汐':>9} | {'有效注册':>8} | {'剔除':>6}") print("-" * 95) mar_start = dates.index('2026-03-01') apr_end = dates.index('2026-04-30') + 1 for i in range(mar_start, apr_end): d = dates[i] dw = date_series[i].strftime('%a') tag = affected_reason.get(d, '—') trend_val = trend_raw[i] tdow_val = trend_with_dow[i] eff_val = effective[i] removed = values[i] - eff_val print(f"{d:>12} {dw:>4} | {values[i]:>5.0f} | {tag:>10} | {trend_val:>6.0f} | {tdow_val:>9.0f} | {eff_val:>8.0f} | {removed:>+6.0f}") # ==================== 汇总 ==================== mar_apr_actual = np.sum(values[mar_start:apr_end]) mar_apr_eff = np.sum(effective[mar_start:apr_end]) mar_apr_affected_count = int(np.sum(is_affected[mar_start:apr_end])) print(f"\n{'='*80}") print("📊 最终汇总") print(f"{'='*80}") print(f" 统计天数: {apr_end - mar_start}") print(f" 活动影响天数: {mar_apr_affected_count}") print(f"") print(f" 实际注册总计: {mar_apr_actual:.0f}") print(f" 有效注册总计: {mar_apr_eff:.0f}") print(f" 剔除活动量: {mar_apr_actual - mar_apr_eff:.0f} 人 ({(mar_apr_actual-mar_apr_eff)/mar_apr_actual*100:.1f}%)") print(f"") print(f" 实际日均: {np.mean(values[mar_start:apr_end]):.0f}") print(f" 有效日均: {np.mean(effective[mar_start:apr_end]):.0f}") print(f"") print(f" 1-2月干净日均: {np.mean(values[:mar_start][clean_mask[:mar_start]]):.0f}") print(f" 3-4月有效日均: {np.mean(effective[mar_start:apr_end]):.0f}") # ==================== 对比 V3 固定值 vs V4 比例 ==================== print(f"\n{'='*80}") print("📊 V3(固定补偿) vs V4(比例补偿) 对比") print(f"{'='*80}") # 用固定值方式算一遍做对比 residuals_abs = clean_y - trend_raw[clean_mask] dow_abs = np.zeros(7) for d in range(7): mask = clean_dow == d if np.sum(mask) > 0: dow_abs[d] = np.mean(residuals_abs[mask]) dow_abs -= np.mean(dow_abs) trend_v3 = trend_raw + dow_abs[dow] trend_v3 = np.maximum(trend_v3, 0) eff_v3 = np.where(is_affected, np.minimum(values, trend_v3), values) print(f"{'方法':>20} | {'有效总计':>10} | {'剔除':>8} | {'有效日均':>8}") print(f"{'V3 固定值补偿':>20} | {np.sum(eff_v3[mar_start:apr_end]):>10.0f} | {mar_apr_actual - np.sum(eff_v3[mar_start:apr_end]):>8.0f} | {np.mean(eff_v3[mar_start:apr_end]):>8.1f}") print(f"{'V4 比例补偿':>20} | {mar_apr_eff:>10.0f} | {mar_apr_actual - mar_apr_eff:>8.0f} | {np.mean(effective[mar_start:apr_end]):>8.1f}") # 列出差异较大的日期 print(f"\n V3 vs V4 有效注册差异 > 3 的日期：") diff = effective[mar_start:apr_end] - eff_v3[mar_start:apr_end] for i in range(len(diff)): if abs(diff[i]) > 3: d = dates[mar_start + i] sign = '+' if diff[i] > 0 else '' print(f" {d}: V3={eff_v3[mar_start+i]:.0f}, V4={effective[mar_start+i]:.0f}, 差异={sign}{diff[i]:.0f}") # ==================== 导出Excel ==================== rows = [] for i in range(mar_start, apr_end): rows.append({ '日期': dates[i], '星期': date_series[i].strftime('%A'), '实际注册人数': int(values[i]), '是否活动影响': '是' if is_affected[i] else '否', '活动说明': affected_reason.get(dates[i], ''), 'LOESS趋势': round(trend_raw[i], 0), '趋势+潮汐(比例)': round(trend_with_dow[i], 0), '有效注册人数(用于转化)': round(effective[i], 0), '剔除人数': round(values[i] - effective[i], 0), }) output_df = pd.DataFrame(rows) output_path = '/root/.openclaw/workspace/output/2026_3-4月注册回归分析.xlsx' with pd.ExcelWriter(output_path, engine='openpyxl') as writer: output_df.to_excel(writer, sheet_name='每日明细', index=False) summary_rows = [ ('统计天数', apr_end - mar_start), ('活动影响天数', mar_apr_affected_count), ('', ''), ('实际注册总计', f'{mar_apr_actual:.0f}'), ('有效注册总计（用于转化）', f'{mar_apr_eff:.0f}'), ('剔除活动量', f'{mar_apr_actual - mar_apr_eff:.0f}'), ('剔除比例', f'{(mar_apr_actual-mar_apr_eff)/mar_apr_actual*100:.1f}%'), ('', ''), ('实际日均', f'{np.mean(values[mar_start:apr_end]):.1f}'), ('有效日均', f'{np.mean(effective[mar_start:apr_end]):.1f}'), ('', ''), ('1-2月干净数据日均', f'{np.mean(values[:mar_start][clean_mask[:mar_start]]):.1f}'), ('3-4月有效数据日均', f'{np.mean(effective[mar_start:apr_end]):.1f}'), ('', ''), ('星期效应-周一', f'{dow_ratio[0]*100:+.1f}%'), ('星期效应-周二', f'{dow_ratio[1]*100:+.1f}%'), ('星期效应-周三', f'{dow_ratio[2]*100:+.1f}%'), ('星期效应-周四', f'{dow_ratio[3]*100:+.1f}%'), ('星期效应-周五', f'{dow_ratio[4]*100:+.1f}%'), ('星期效应-周六', f'{dow_ratio[5]*100:+.1f}%'), ('星期效应-周日', f'{dow_ratio[6]*100:+.1f}%'), ] pd.DataFrame(summary_rows, columns=['指标','值']).to_excel(writer, sheet_name='汇总', index=False) print(f"\n✅ Excel 已导出: {output_path}")