Add stock prediction experiments: AAPL forecast + 4-version backtest comparison

- stock_forecast.py: Basic AAPL price forecast
- stock_backtest.py: V1 original price prediction backtest
- stock_backtest_optimized.py: V2 log-return + ensemble optimization
- stock_backtest_xreg.py: V3 XReg covariates (volume/RSI/SPY)
- stock_backtest_trick.py: V4 SPY trend guidance + volatility adjustment
- Visualization charts for all versions

stock_backtest.py

@@ -0,0 +1,115 @@
+import yfinance as yf
+import numpy as np
+import torch
+import timesfm
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from datetime import datetime, timedelta
+
+# 1. 获取 AAPL 数据（到今天为止）
+print("下载 AAPL 股票数据...", flush=True)
+end = datetime.now()
+start = end - timedelta(days=365)
+df = yf.download("AAPL", start=start.strftime("%Y-%m-%d"), end=end.strftime("%Y-%m-%d"), progress=False)
+close = df["Close"].values.flatten().astype(np.float32)
+dates = df.index
+
+# 2. 分割：6/15 之前为训练集，6/16 之后为实际值
+split_date = "2026-06-15"
+split_idx = None
+for i, d in enumerate(dates):
+    if str(d.date()) <= split_date:
+        split_idx = i
+# split_idx 是 <= 6/15 的最后一个索引
+train_data = close[: split_idx + 1]
+train_dates = dates[: split_idx + 1]
+actual_data = close[split_idx + 1 :]
+actual_dates = dates[split_idx + 1 :]
+
+HORIZON = len(actual_data)
+print(f"训练数据: {len(train_data)} 天（截至 {train_dates[-1].date()}）", flush=True)
+print(f"实际数据: {HORIZON} 天（{actual_dates[0].date()} ~ {actual_dates[-1].date()}）", flush=True)
+print(f"分割日收盘价: {train_data[-1]:.2f}", flush=True)
+
+# 3. 加载模型
+print()
+print("加载 TimesFM 模型...", flush=True)
+torch.set_float32_matmul_precision("high")
+model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
+    "google/timesfm-2.5-200m-pytorch", torch_compile=False
+)
+print("模型加载完成", flush=True)
+
+# 4. 编译并预测
+model.compile(
+    timesfm.ForecastConfig(
+        max_context=512,
+        max_horizon=128,
+        normalize_inputs=True,
+        use_continuous_quantile_head=True,
+        force_flip_invariance=False,
+        infer_is_positive=True,
+        fix_quantile_crossing=True,
+    )
+)
+print(f"预测未来 {HORIZON} 个交易日...", flush=True)
+pf, qf = model.forecast(horizon=HORIZON, inputs=[train_data])
+print("预测完成！", flush=True)
+
+# 5. 计算误差指标
+actual = actual_data
+pred = pf[0][:HORIZON]
+mae = np.mean(np.abs(actual - pred))
+rmse = np.sqrt(np.mean((actual - pred) ** 2))
+mape = np.mean(np.abs((actual - pred) / actual)) * 100
+
+# 方向准确率
+actual_dir = np.diff(actual)
+pred_dir = np.diff(pred)
+dir_acc = np.mean(actual_dir * pred_dir > 0) * 100
+
+# 置信区间覆盖率
+in_80 = np.mean((actual >= qf[0, :HORIZON, 1]) & (actual <= qf[0, :HORIZON, 9])) * 100
+in_40 = np.mean((actual >= qf[0, :HORIZON, 3]) & (actual <= qf[0, :HORIZON, 7])) * 100
+
+print()
+print("=== 回测结果：6/16 ~ 6/30 预测 vs 实际 ===", flush=True)
+print()
+print("  日期          实际价    预测价    误差    误差%", flush=True)
+for i in range(HORIZON):
+    err = pred[i] - actual[i]
+    err_pct = err / actual[i] * 100
+    print(
+        f"  {actual_dates[i].date()}  {actual[i]:7.2f}  {pred[i]:7.2f}  {err:+7.2f}  {err_pct:+6.2f}%",
+        flush=True,
+    )
+
+print()
+print("=== 误差指标 ===", flush=True)
+print(f"  MAE  (平均绝对误差):     ${mae:.2f}", flush=True)
+print(f"  RMSE (均方根误差):       ${rmse:.2f}", flush=True)
+print(f"  MAPE (平均绝对百分比误差): {mape:.2f}%", flush=True)
+print(f"  方向准确率:               {dir_acc:.1f}%", flush=True)
+print(f"  80%置信区间覆盖率:        {in_80:.1f}%", flush=True)
+print(f"  40%置信区间覆盖率:        {in_40:.1f}%", flush=True)
+
+# 6. 画图
+fig, ax = plt.subplots(figsize=(14, 6))
+show_n = min(60, len(train_data))
+ax.plot(range(show_n), train_data[-show_n:], label="Historical Close", color="steelblue", linewidth=1.5)
+x_actual = range(show_n, show_n + HORIZON)
+ax.plot(x_actual, actual, label="Actual", color="forestgreen", linewidth=2, marker="o", markersize=3)
+ax.plot(x_actual, pred, label="Forecast", color="tomato", linewidth=2, linestyle="--")
+ax.fill_between(x_actual, qf[0, :HORIZON, 1], qf[0, :HORIZON, 9], alpha=0.15, color="tomato", label="80% CI")
+ax.fill_between(x_actual, qf[0, :HORIZON, 3], qf[0, :HORIZON, 7], alpha=0.3, color="tomato", label="40% CI")
+ax.axvline(x=show_n - 1, color="gray", linestyle="--", alpha=0.5, label="Forecast Start")
+ax.set_title(f"AAPL Backtest: Forecast vs Actual (MAPE={mape:.2f}%, Dir.Acc={dir_acc:.0f}%)", fontsize=14)
+ax.set_xlabel("Trading Days")
+ax.set_ylabel("Price (USD)")
+ax.legend(loc="upper left")
+plt.tight_layout()
+plt.savefig("aapl_backtest.png", dpi=150)
+print()
+print("Chart saved: aapl_backtest.png", flush=True)
+print("Done!", flush=True)

stock_backtest_optimized.py

@@ -0,0 +1,184 @@
+import yfinance as yf
+import numpy as np
+import torch
+import timesfm
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from datetime import datetime, timedelta
+
+# ---------- 技术指标计算 ----------
+def calc_rsi(prices, period=14):
+    deltas = np.diff(prices)
+    gains = np.where(deltas > 0, deltas, 0.0)
+    losses = np.where(deltas < 0, -deltas, 0.0)
+    avg_gain = np.convolve(gains, np.ones(period) / period, mode="valid")
+    avg_loss = np.convolve(losses, np.ones(period) / period, mode="valid")
+    avg_loss = np.where(avg_loss == 0, 1e-10, avg_loss)
+    rs = avg_gain / avg_loss
+    rsi = 100.0 - (100.0 / (1.0 + rs))
+    # pad front to match length
+    return np.concatenate([np.full(period, 50.0), rsi])
+
+
+def calc_sma(prices, period):
+    sma = np.convolve(prices, np.ones(period) / period, mode="valid")
+    return np.concatenate([np.full(period - 1, sma[0] if len(sma) > 0 else 0.0), sma])
+
+
+# ---------- 1. 获取数据 ----------
+print("下载 AAPL 股票数据...", flush=True)
+end = datetime.now()
+start = end - timedelta(days=365)
+df = yf.download("AAPL", start=start.strftime("%Y-%m-%d"), end=end.strftime("%Y-%m-%d"), progress=False)
+close = df["Close"].values.flatten().astype(np.float32)
+volume = df["Volume"].values.flatten().astype(np.float32)
+dates = df.index
+
+# 分割
+split_date = "2026-06-15"
+split_idx = None
+for i, d in enumerate(dates):
+    if str(d.date()) <= split_date:
+        split_idx = i
+
+train_close = close[: split_idx + 1]
+train_vol = volume[: split_idx + 1]
+train_dates = dates[: split_idx + 1]
+actual_close = close[split_idx + 1 :]
+actual_dates = dates[split_idx + 1 :]
+HORIZON = len(actual_close)
+
+print(f"训练数据: {len(train_close)} 天（截至 {train_dates[-1].date()}）", flush=True)
+print(f"实际数据: {HORIZON} 天（{actual_dates[0].date()} ~ {actual_dates[-1].date()}）", flush=True)
+
+# ---------- 2. 计算对数收益率 ----------
+# log returns = ln(P_t / P_{t-1})，更平稳
+train_logret = np.diff(np.log(train_close)).astype(np.float32)  # length = N-1
+actual_logret = np.diff(np.log(np.concatenate([train_close[-1:], actual_close]))).astype(np.float32)
+
+print(f"对数收益率: 均值={train_logret.mean():.6f}, 标准差={train_logret.std():.6f}", flush=True)
+
+# ---------- 3. 加载模型 ----------
+print()
+print("加载 TimesFM 模型...", flush=True)
+torch.set_float32_matmul_precision("high")
+model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
+    "google/timesfm-2.5-200m-pytorch", torch_compile=False
+)
+print("模型加载完成", flush=True)
+
+# ---------- 4. 多窗口集成预测（对数收益率） ----------
+CONTEXTS = [128, 256, 512]  # 不同 context 长度
+all_preds = []
+all_quantiles = []
+
+for ctx_len in CONTEXTS:
+    ctx_data = train_logret[-ctx_len:] if len(train_logret) >= ctx_len else train_logret
+    actual_ctx = min(ctx_len, len(ctx_data))
+    # 对收益率: infer_is_positive=False (可负), normalize_inputs=True
+    model.compile(
+        timesfm.ForecastConfig(
+            max_context=actual_ctx,
+            max_horizon=128,
+            normalize_inputs=True,
+            use_continuous_quantile_head=True,
+            force_flip_invariance=False,
+            infer_is_positive=False,  # 收益率可正可负
+            fix_quantile_crossing=True,
+        )
+    )
+    pf, qf = model.forecast(horizon=HORIZON, inputs=[ctx_data])
+    all_preds.append(pf[0][:HORIZON])
+    all_quantiles.append(qf[0][:HORIZON])
+    print(f"  context={actual_ctx} 预测完成", flush=True)
+
+# 集成: 取平均
+ensemble_pred_logret = np.mean(all_preds, axis=0)
+ensemble_q_logret = np.mean(all_quantiles, axis=0)
+
+# ---------- 5. 转换回价格 ----------
+# P_t = P_{t-1} * exp(r_t)
+last_price = train_close[-1]
+pred_prices = []
+for i in range(HORIZON):
+    last_price = last_price * np.exp(ensemble_pred_logret[i])
+    pred_prices.append(last_price)
+pred_prices = np.array(pred_prices)
+
+# 分位数价格
+q_prices = np.zeros((HORIZON, 10))
+for qi in range(10):
+    p = train_close[-1]
+    for i in range(HORIZON):
+        p = p * np.exp(ensemble_q_logret[i, qi])
+        q_prices[i, qi] = p
+
+# ---------- 6. 计算误差 ----------
+actual = actual_close
+pred = pred_prices
+mae = np.mean(np.abs(actual - pred))
+rmse = np.sqrt(np.mean((actual - pred) ** 2))
+mape = np.mean(np.abs((actual - pred) / actual)) * 100
+
+actual_dir = np.diff(actual)
+pred_dir = np.diff(pred)
+dir_acc = np.mean(actual_dir * pred_dir > 0) * 100
+
+in_80 = np.mean((actual >= q_prices[:, 1]) & (actual <= q_prices[:, 9])) * 100
+in_40 = np.mean((actual >= q_prices[:, 3]) & (actual <= q_prices[:, 7])) * 100
+
+print()
+print("=== 优化版回测结果：6/16 ~ 6/30 ===", flush=True)
+print()
+print("  日期          实际价    预测价    误差    误差%", flush=True)
+for i in range(HORIZON):
+    err = pred[i] - actual[i]
+    err_pct = err / actual[i] * 100
+    print(
+        f"  {actual_dates[i].date()}  {actual[i]:7.2f}  {pred[i]:7.2f}  {err:+7.2f}  {err_pct:+6.2f}%",
+        flush=True,
+    )
+
+print()
+print("=== 误差指标（优化版 vs 原始版）===", flush=True)
+print(f"  MAE:    ${mae:.2f}        (原始: $8.46)", flush=True)
+print(f"  RMSE:   ${rmse:.2f}       (原始: $11.29)", flush=True)
+print(f"  MAPE:   {mape:.2f}%       (原始: 2.98%)", flush=True)
+print(f"  方向准确率: {dir_acc:.1f}%       (原始: 55.6%)", flush=True)
+print(f"  80%CI覆盖率: {in_80:.1f}%        (原始: 90.0%)", flush=True)
+print(f"  40%CI覆盖率: {in_40:.1f}%        (原始: 60.0%)", flush=True)
+
+# ---------- 7. 画对比图 ----------
+fig, axes = plt.subplots(2, 1, figsize=(14, 10), sharex=False)
+
+# 上图: 价格对比
+ax = axes[0]
+show_n = min(60, len(train_close))
+ax.plot(range(show_n), train_close[-show_n:], label="Historical", color="steelblue", linewidth=1.5)
+x_actual = range(show_n, show_n + HORIZON)
+ax.plot(x_actual, actual, label="Actual", color="forestgreen", linewidth=2, marker="o", markersize=4)
+ax.plot(x_actual, pred, label="Optimized Forecast", color="tomato", linewidth=2, linestyle="--")
+ax.fill_between(x_actual, q_prices[:, 1], q_prices[:, 9], alpha=0.15, color="tomato", label="80% CI")
+ax.fill_between(x_actual, q_prices[:, 3], q_prices[:, 7], alpha=0.3, color="tomato", label="40% CI")
+ax.axvline(x=show_n - 1, color="gray", linestyle="--", alpha=0.5)
+ax.set_title(f"Optimized: Log-Return + Ensemble (MAPE={mape:.2f}%, Dir={dir_acc:.0f}%)", fontsize=13)
+ax.set_ylabel("Price (USD)")
+ax.legend(loc="upper left")
+
+# 下图: 预测误差对比
+ax2 = axes[1]
+errors = pred - actual
+ax2.bar(range(HORIZON), errors, color=["tomato" if e > 0 else "steelblue" for e in errors], alpha=0.7)
+ax2.axhline(y=0, color="black", linewidth=0.8)
+ax2.set_title("Forecast Error per Day (Optimized)", fontsize=13)
+ax2.set_xlabel("Trading Days After Split")
+ax2.set_ylabel("Error (USD)")
+ax2.set_xticks(range(HORIZON))
+ax2.set_xticklabels([str(d.date()) for d in actual_dates], rotation=45, fontsize=8)
+
+plt.tight_layout()
+plt.savefig("aapl_backtest_optimized.png", dpi=150)
+print()
+print("Chart saved: aapl_backtest_optimized.png", flush=True)
+print("Done!", flush=True)

stock_backtest_trick.py

@@ -0,0 +1,217 @@
+"""TimesFM 纯技巧优化：
+1. 对数收益率预测（更平稳）
+2. 多窗口集成（128/256/512）
+3. SPY 大盘走势引导：先预测 SPY，用 SPY 预测的趋势辅助判断 AAPL 方向
+4. 波动率调整：用近期波动率缩放置信区间
+5. infer_is_positive=False（收益率可负）
+"""
+import yfinance as yf
+import numpy as np
+import torch
+import timesfm
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from datetime import datetime, timedelta
+
+
+# 1. 获取数据
+print("下载 AAPL + SPY 数据...", flush=True)
+end = datetime.now()
+start = end - timedelta(days=365)
+df_aapl = yf.download("AAPL", start=start.strftime("%Y-%m-%d"), end=end.strftime("%Y-%m-%d"), progress=False)
+df_spy = yf.download("SPY", start=start.strftime("%Y-%m-%d"), end=end.strftime("%Y-%m-%d"), progress=False)
+
+close = df_aapl["Close"].values.flatten().astype(np.float32)
+spy_close = df_spy["Close"].values.flatten().astype(np.float32)
+dates = df_aapl.index
+
+min_len = min(len(close), len(spy_close))
+close = close[-min_len:]
+spy_close = spy_close[-min_len:]
+dates = dates[-min_len:]
+
+# 分割
+split_date = "2026-06-15"
+split_idx = None
+for i, d in enumerate(dates):
+    if str(d.date()) <= split_date:
+        split_idx = i
+
+train_close = close[: split_idx + 1]
+train_spy = spy_close[: split_idx + 1]
+train_dates = dates[: split_idx + 1]
+actual_close = close[split_idx + 1 :]
+actual_spy = spy_close[split_idx + 1 :]
+actual_dates = dates[split_idx + 1 :]
+HORIZON = len(actual_close)
+
+print(f"训练数据: {len(train_close)} 天 | 预测: {HORIZON} 天", flush=True)
+
+# 2. 对数收益率
+train_logret = np.diff(np.log(train_close)).astype(np.float32)
+train_spy_logret = np.diff(np.log(train_spy)).astype(np.float32)
+actual_logret = np.diff(np.log(np.concatenate([train_close[-1:], actual_close]))).astype(np.float32)
+
+# 3. 加载模型
+print()
+print("加载 TimesFM 模型...", flush=True)
+torch.set_float32_matmul_precision("high")
+model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
+    "google/timesfm-2.5-200m-pytorch", torch_compile=False
+)
+print("模型加载完成", flush=True)
+
+# 4. 多窗口集成预测 AAPL 收益率
+CONTEXTS = [128, 256, 512]
+all_preds = []
+all_quantiles = []
+
+for ctx_len in CONTEXTS:
+    ctx_data = train_logret[-ctx_len:] if len(train_logret) >= ctx_len else train_logret
+    actual_ctx = min(ctx_len, len(ctx_data))
+    # 向上取整到 32 的倍数
+    actual_ctx = ((actual_ctx + 31) // 32) * 32
+    ctx_data = train_logret[-actual_ctx:]
+
+    model.compile(
+        timesfm.ForecastConfig(
+            max_context=actual_ctx,
+            max_horizon=128,
+            normalize_inputs=True,
+            use_continuous_quantile_head=True,
+            force_flip_invariance=False,
+            infer_is_positive=False,
+            fix_quantile_crossing=True,
+        )
+    )
+    pf, qf = model.forecast(horizon=HORIZON, inputs=[ctx_data])
+    all_preds.append(pf[0][:HORIZON])
+    all_quantiles.append(qf[0][:HORIZON])
+    print(f"  AAPL context={actual_ctx} done", flush=True)
+
+# 5. 预测 SPY 收益率（大盘趋势引导）
+spy_ctx = train_spy_logret[-256:]
+spy_ctx_len = ((len(spy_ctx) + 31) // 32) * 32
+spy_ctx = train_spy_logret[-spy_ctx_len:]
+model.compile(
+    timesfm.ForecastConfig(
+        max_context=spy_ctx_len,
+        max_horizon=128,
+        normalize_inputs=True,
+        use_continuous_quantile_head=True,
+        force_flip_invariance=False,
+        infer_is_positive=False,
+        fix_quantile_crossing=True,
+    )
+)
+spy_pf, _ = model.forecast(horizon=HORIZON, inputs=[spy_ctx])
+print(f"  SPY context={spy_ctx_len} done", flush=True)
+
+# 6. 集成 + SPY 趋势调整
+ensemble_pred_logret = np.mean(all_preds, axis=0)
+ensemble_q_logret = np.mean(all_quantiles, axis=0)
+spy_pred_logret = spy_pf[0][:HORIZON]
+
+# SPY 趋势调整：如果 SPY 预测下跌，对 AAPL 预测施加向下的调整
+# 计算 AAPL 对 SPY 的 beta（敏感度）
+beta = np.corrcoef(train_logret[-60:], train_spy_logret[-60:])[0, 1]
+print(f"  AAPL-SPY 60日相关系数: {beta:.3f}", flush=True)
+
+# 调整：将 SPY 预测的偏离均值部分 * beta 加到 AAPL 预测上
+spy_mean = np.mean(train_spy_logret[-60:])
+spy_deviation = spy_pred_logret - spy_mean  # SPY 偏离其均值的部分
+adjustment = beta * spy_deviation * 0.3  # 0.3 是调整强度，避免过度修正
+adjusted_pred_logret = ensemble_pred_logret + adjustment
+
+# 7. 波动率调整置信区间
+recent_vol = np.std(train_logret[-20:])
+long_vol = np.std(train_logret[-60:])
+print(f"  近20日波动率: {recent_vol:.5f} | 近60日波动率: {long_vol:.5f}", flush=True)
+
+# 如果近期波动率高于长期，扩大置信区间
+vol_ratio = recent_vol / max(long_vol, 1e-8)
+vol_scale = max(vol_ratio, 1.0)  # 只扩大不缩小
+adjusted_q_logret = ensemble_q_logret.copy()
+median_idx = 5
+for qi in range(10):
+    if qi != median_idx:
+        adjusted_q_logret[:, qi] = ensemble_q_logret[:, median_idx] + (
+            ensemble_q_logret[:, qi] - ensemble_q_logret[:, median_idx]
+        ) * vol_scale
+
+# 8. 转换回价格
+last_price = train_close[-1]
+pred_prices = []
+for i in range(HORIZON):
+    last_price = last_price * np.exp(adjusted_pred_logret[i])
+    pred_prices.append(last_price)
+pred_prices = np.array(pred_prices)
+
+q_prices = np.zeros((HORIZON, 10))
+for qi in range(10):
+    p = train_close[-1]
+    for i in range(HORIZON):
+        p = p * np.exp(adjusted_q_logret[i, qi])
+        q_prices[i, qi] = p
+
+# 9. 误差指标
+actual = actual_close
+pred = pred_prices
+mae = np.mean(np.abs(actual - pred))
+rmse = np.sqrt(np.mean((actual - pred) ** 2))
+mape = np.mean(np.abs((actual - pred) / actual)) * 100
+actual_dir = np.diff(actual)
+pred_dir = np.diff(pred)
+dir_acc = np.mean(actual_dir * pred_dir > 0) * 100
+in_80 = np.mean((actual >= q_prices[:, 1]) & (actual <= q_prices[:, 9])) * 100
+in_40 = np.mean((actual >= q_prices[:, 3]) & (actual <= q_prices[:, 7])) * 100
+
+print()
+print("=== 纯技巧优化版回测结果 ===", flush=True)
+print()
+print("  日期          实际价    预测价    误差    误差%", flush=True)
+for i in range(HORIZON):
+    err = pred[i] - actual[i]
+    err_pct = err / actual[i] * 100
+    print(f"  {actual_dates[i].date()}  {actual[i]:7.2f}  {pred[i]:7.2f}  {err:+7.2f}  {err_pct:+6.2f}%", flush=True)
+
+print()
+print("=== 误差指标（四版对比）===", flush=True)
+print(f"  MAE:    ${mae:.2f}        (原始: $8.46, 对数收益率: $8.00, XReg: $9.21)", flush=True)
+print(f"  RMSE:   ${rmse:.2f}       (原始: $11.29, 对数收益率: $11.09, XReg: $12.47)", flush=True)
+print(f"  MAPE:   {mape:.2f}%       (原始: 2.98%, 对数收益率: 2.82%, XReg: 3.25%)", flush=True)
+print(f"  方向准确率: {dir_acc:.1f}%       (原始: 55.6%, 对数收益率: 44.4%, XReg: 44.4%)", flush=True)
+print(f"  80%CI覆盖率: {in_80:.1f}%        (原始: 90.0%, 对数收益率: 100.0%, XReg: 70.0%)", flush=True)
+print(f"  40%CI覆盖率: {in_40:.1f}%        (原始: 60.0%, 对数收益率: 80.0%, XReg: 40.0%)", flush=True)
+
+# 10. 画图
+fig, axes = plt.subplots(2, 1, figsize=(14, 10))
+show_n = min(60, len(train_close))
+ax = axes[0]
+ax.plot(range(show_n), train_close[-show_n:], label="Historical", color="steelblue", linewidth=1.5)
+x_actual = range(show_n, show_n + HORIZON)
+ax.plot(x_actual, actual, label="Actual", color="forestgreen", linewidth=2, marker="o", markersize=4)
+ax.plot(x_actual, pred, label="Trick Forecast", color="purple", linewidth=2, linestyle="--")
+ax.fill_between(x_actual, q_prices[:, 1], q_prices[:, 9], alpha=0.15, color="purple", label="80% CI")
+ax.fill_between(x_actual, q_prices[:, 3], q_prices[:, 7], alpha=0.3, color="purple", label="40% CI")
+ax.axvline(x=show_n - 1, color="gray", linestyle="--", alpha=0.5)
+ax.set_title(f"Trick: LogRet + Ensemble + SPY + VolAdj (MAPE={mape:.2f}%, Dir={dir_acc:.0f}%)", fontsize=13)
+ax.set_ylabel("Price (USD)")
+ax.legend(loc="upper left")
+
+ax2 = axes[1]
+errors = pred - actual
+ax2.bar(range(HORIZON), errors, color=["purple" if e > 0 else "steelblue" for e in errors], alpha=0.7)
+ax2.axhline(y=0, color="black", linewidth=0.8)
+ax2.set_title("Trick Forecast Error per Day", fontsize=13)
+ax2.set_xlabel("Trading Days After Split")
+ax2.set_ylabel("Error (USD)")
+ax2.set_xticks(range(HORIZON))
+ax2.set_xticklabels([str(d.date()) for d in actual_dates], rotation=45, fontsize=8)
+
+plt.tight_layout()
+plt.savefig("aapl_backtest_trick.png", dpi=150)
+print()
+print("Chart saved: aapl_backtest_trick.png", flush=True)
+print("Done!", flush=True)

stock_backtest_xreg.py

@@ -0,0 +1,203 @@
+"""TimesFM + XReg 协变量预测：加入成交量、RSI、SPY 大盘指数"""
+import yfinance as yf
+import numpy as np
+import torch
+import timesfm
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from datetime import datetime, timedelta
+
+
+def calc_rsi(prices, period=14):
+    deltas = np.diff(prices)
+    gains = np.where(deltas > 0, deltas, 0.0)
+    losses = np.where(deltas < 0, -deltas, 0.0)
+    avg_gain = np.convolve(gains, np.ones(period) / period, mode="valid")
+    avg_loss = np.convolve(losses, np.ones(period) / period, mode="valid")
+    avg_loss = np.where(avg_loss == 0, 1e-10, avg_loss)
+    rs = avg_gain / avg_loss
+    rsi = 100.0 - (100.0 / (1.0 + rs))
+    return np.concatenate([np.full(period, 50.0), rsi])
+
+
+# 1. 获取数据
+print("下载 AAPL + SPY 数据...", flush=True)
+end = datetime.now()
+start = end - timedelta(days=365)
+df_aapl = yf.download("AAPL", start=start.strftime("%Y-%m-%d"), end=end.strftime("%Y-%m-%d"), progress=False)
+df_spy = yf.download("SPY", start=start.strftime("%Y-%m-%d"), end=end.strftime("%Y-%m-%d"), progress=False)
+
+close = df_aapl["Close"].values.flatten().astype(np.float64)
+volume = df_aapl["Volume"].values.flatten().astype(np.float64)
+spy_close = df_spy["Close"].values.flatten().astype(np.float64)
+dates = df_aapl.index
+
+# 对齐长度
+min_len = min(len(close), len(spy_close))
+close = close[-min_len:]
+volume = volume[-min_len:]
+spy_close = spy_close[-min_len:]
+dates = dates[-min_len:]
+
+# 计算 RSI
+rsi = calc_rsi(close)
+
+# 分割
+split_date = "2026-06-15"
+split_idx = None
+for i, d in enumerate(dates):
+    if str(d.date()) <= split_date:
+        split_idx = i
+
+train_close = close[: split_idx + 1]
+train_vol = volume[: split_idx + 1]
+train_spy = spy_close[: split_idx + 1]
+train_rsi = rsi[: split_idx + 1]
+train_dates = dates[: split_idx + 1]
+
+actual_close = close[split_idx + 1 :]
+actual_dates = dates[split_idx + 1 :]
+HORIZON = len(actual_close)
+
+print(f"训练数据: {len(train_close)} 天", flush=True)
+print(f"预测目标: {HORIZON} 天", flush=True)
+
+# 2. 准备协变量
+# 动态数值协变量需要覆盖 context + horizon 的完整长度
+# 对每个序列: train 部分用实际值，test 部分需要"未来值"
+# 对于 RSI 和 Volume，我们没有未来值，用最后一个值填充
+# 对于 SPY，用训练集最后一个值填充（因为我们无法预知未来 SPY）
+
+# 协变量需要: 每个协变量是一个 list，每个元素对应一个输入序列的完整长度(context+horizon)
+full_len = len(train_close) + HORIZON
+
+# 成交量: train 用实际值，future 用最近 5 日均值
+vol_future = np.mean(train_vol[-5:])
+vol_full = np.concatenate([train_vol, np.full(HORIZON, vol_future)])
+
+# RSI: train 用实际值，future 用 50（中性）
+rsi_full = np.concatenate([train_rsi, np.full(HORIZON, 50.0)])
+
+# SPY: train 用实际值，future 用最后一个值
+spy_future = train_spy[-1]
+spy_full = np.concatenate([train_spy, np.full(HORIZON, spy_future)])
+
+# 3. 加载模型
+print()
+print("加载 TimesFM 模型...", flush=True)
+torch.set_float32_matmul_precision("high")
+model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
+    "google/timesfm-2.5-200m-pytorch", torch_compile=False
+)
+print("模型加载完成", flush=True)
+
+# 4. 编译（return_backcast=True 是 XReg 必需的）
+model.compile(
+    timesfm.ForecastConfig(
+        max_context=512,
+        max_horizon=128,
+        normalize_inputs=True,
+        use_continuous_quantile_head=True,
+        force_flip_invariance=False,
+        infer_is_positive=True,
+        fix_quantile_crossing=True,
+        return_backcast=True,  # XReg 需要
+    )
+)
+print("编译完成", flush=True)
+
+# 5. 用 XReg 预测
+print("运行 XReg 协变量预测...", flush=True)
+
+# 动态数值协变量: dict[str, list[list[float]]]
+# 每个协变量是一个 list，其中每个元素是一个序列（对应一个输入）
+# 这里只有一个输入序列
+dynamic_num_covs = {
+    "volume": [vol_full],
+    "rsi": [rsi_full],
+    "spy_close": [spy_full],
+}
+
+# 静态数值协变量
+static_num_covs = {
+    "avg_volume": [np.mean(train_vol)],
+}
+
+point_outputs, quantile_outputs = model.forecast_with_covariates(
+    inputs=[train_close],
+    dynamic_numerical_covariates=dynamic_num_covs,
+    static_numerical_covariates=static_num_covs,
+    xreg_mode="xreg + timesfm",  # 先回归再预测残差
+    normalize_xreg_target_per_input=True,
+    ridge=1.0,
+)
+print("XReg 预测完成！", flush=True)
+
+# 6. 提取结果
+pred = np.array(point_outputs[0][:HORIZON])
+q = np.array(quantile_outputs[0])  # (horizon, 10) or (full, 10)
+# quantile_outputs 可能包含 backcast，取最后 HORIZON 个
+if q.shape[0] > HORIZON:
+    q = q[-HORIZON:]
+
+# 7. 计算误差
+actual = actual_close
+mae = np.mean(np.abs(actual - pred))
+rmse = np.sqrt(np.mean((actual - pred) ** 2))
+mape = np.mean(np.abs((actual - pred) / actual)) * 100
+actual_dir = np.diff(actual)
+pred_dir = np.diff(pred)
+dir_acc = np.mean(actual_dir * pred_dir > 0) * 100
+in_80 = np.mean((actual >= q[:, 1]) & (actual <= q[:, 9])) * 100
+in_40 = np.mean((actual >= q[:, 3]) & (actual <= q[:, 7])) * 100
+
+print()
+print("=== XReg 协变量版回测结果 ===", flush=True)
+print()
+print("  日期          实际价    预测价    误差    误差%", flush=True)
+for i in range(HORIZON):
+    err = pred[i] - actual[i]
+    err_pct = err / actual[i] * 100
+    print(f"  {actual_dates[i].date()}  {actual[i]:7.2f}  {pred[i]:7.2f}  {err:+7.2f}  {err_pct:+6.2f}%", flush=True)
+
+print()
+print("=== 误差指标（三版对比）===", flush=True)
+print(f"  MAE:    ${mae:.2f}        (原始: $8.46, 优化: $8.00)", flush=True)
+print(f"  RMSE:   ${rmse:.2f}       (原始: $11.29, 优化: $11.09)", flush=True)
+print(f"  MAPE:   {mape:.2f}%       (原始: 2.98%, 优化: 2.82%)", flush=True)
+print(f"  方向准确率: {dir_acc:.1f}%       (原始: 55.6%, 优化: 44.4%)", flush=True)
+print(f"  80%CI覆盖率: {in_80:.1f}%        (原始: 90.0%, 优化: 100.0%)", flush=True)
+print(f"  40%CI覆盖率: {in_40:.1f}%        (原始: 60.0%, 优化: 80.0%)", flush=True)
+
+# 8. 画图
+fig, axes = plt.subplots(2, 1, figsize=(14, 10))
+show_n = min(60, len(train_close))
+ax = axes[0]
+ax.plot(range(show_n), train_close[-show_n:], label="Historical", color="steelblue", linewidth=1.5)
+x_actual = range(show_n, show_n + HORIZON)
+ax.plot(x_actual, actual, label="Actual", color="forestgreen", linewidth=2, marker="o", markersize=4)
+ax.plot(x_actual, pred, label="XReg Forecast", color="darkorange", linewidth=2, linestyle="--")
+ax.fill_between(x_actual, q[:, 1], q[:, 9], alpha=0.15, color="darkorange", label="80% CI")
+ax.fill_between(x_actual, q[:, 3], q[:, 7], alpha=0.3, color="darkorange", label="40% CI")
+ax.axvline(x=show_n - 1, color="gray", linestyle="--", alpha=0.5)
+ax.set_title(f"XReg: AAPL + Volume/RSI/SPY (MAPE={mape:.2f}%, Dir={dir_acc:.0f}%)", fontsize=13)
+ax.set_ylabel("Price (USD)")
+ax.legend(loc="upper left")
+
+# 误差柱状图
+ax2 = axes[1]
+errors = pred - actual
+ax2.bar(range(HORIZON), errors, color=["darkorange" if e > 0 else "steelblue" for e in errors], alpha=0.7)
+ax2.axhline(y=0, color="black", linewidth=0.8)
+ax2.set_title("XReg Forecast Error per Day", fontsize=13)
+ax2.set_xlabel("Trading Days After Split")
+ax2.set_ylabel("Error (USD)")
+ax2.set_xticks(range(HORIZON))
+ax2.set_xticklabels([str(d.date()) for d in actual_dates], rotation=45, fontsize=8)
+
+plt.tight_layout()
+plt.savefig("aapl_backtest_xreg.png", dpi=150)
+print()
+print("Chart saved: aapl_backtest_xreg.png", flush=True)
+print("Done!", flush=True)

stock_forecast.py

@@ -0,0 +1,84 @@
+import yfinance as yf
+import numpy as np
+import torch
+import timesfm
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from datetime import datetime, timedelta
+
+# 1. 获取 AAPL 近 1 年收盘价
+print("下载 AAPL 股票数据...", flush=True)
+end = datetime.now()
+start = end - timedelta(days=365)
+df = yf.download("AAPL", start=start.strftime("%Y-%m-%d"), end=end.strftime("%Y-%m-%d"), progress=False)
+close = df["Close"].values.flatten().astype(np.float32)
+dates = df.index
+print(f"获取到 {len(close)} 个交易日", flush=True)
+print(f"价格范围: {close.min():.2f} ~ {close.max():.2f}", flush=True)
+print(f"最近 5 日收盘价: {close[-5:]}", flush=True)
+
+# 2. 加载 TimesFM 模型
+print()
+print("加载 TimesFM 模型...", flush=True)
+torch.set_float32_matmul_precision("high")
+model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
+    "google/timesfm-2.5-200m-pytorch", torch_compile=False
+)
+print("模型加载完成", flush=True)
+
+# 3. 编译并预测未来 20 个交易日
+HORIZON = 20
+model.compile(
+    timesfm.ForecastConfig(
+        max_context=512,
+        max_horizon=128,
+        normalize_inputs=True,
+        use_continuous_quantile_head=True,
+        force_flip_invariance=False,
+        infer_is_positive=True,
+        fix_quantile_crossing=True,
+    )
+)
+print("编译完成，开始预测...", flush=True)
+
+pf, qf = model.forecast(horizon=HORIZON, inputs=[close])
+print("预测完成！", flush=True)
+
+# 4. 输出结果
+print()
+print("=== AAPL 未来 20 个交易日预测 ===", flush=True)
+print(f"当前价格: {close[-1]:.2f}", flush=True)
+print()
+print("  日期(估计)    点预测    q10(低)   q90(高)", flush=True)
+last_date = dates[-1]
+for i in range(HORIZON):
+    est_date = last_date + timedelta(days=i + 1)
+    print(
+        f"  {est_date.strftime('%Y-%m-%d')}  {pf[0][i]:7.2f}  {qf[0][i,1]:7.2f}  {qf[0][i,9]:7.2f}",
+        flush=True,
+    )
+
+print()
+print(f"预测均价: {pf[0].mean():.2f}", flush=True)
+print(f"预测涨跌: {(pf[0][-1] - close[-1]) / close[-1] * 100:+.2f}%", flush=True)
+print(f"80%置信区间: {qf[0,-1,1]:.2f} ~ {qf[0,-1,9]:.2f}", flush=True)
+
+# 5. 画图
+fig, ax = plt.subplots(figsize=(14, 6))
+show_n = min(60, len(close))
+ax.plot(range(show_n), close[-show_n:], label="历史收盘价", color="steelblue", linewidth=1.5)
+x_fc = range(show_n, show_n + HORIZON)
+ax.plot(x_fc, pf[0], label="点预测(中位数)", color="tomato", linewidth=2)
+ax.fill_between(x_fc, qf[0, :, 1], qf[0, :, 9], alpha=0.2, color="tomato", label="80% 置信区间")
+ax.fill_between(x_fc, qf[0, :, 3], qf[0, :, 7], alpha=0.35, color="tomato", label="40% 置信区间")
+ax.axvline(x=show_n - 1, color="gray", linestyle="--", alpha=0.5, label="预测起点")
+ax.set_title("AAPL 收盘价预测 (TimesFM 2.5)", fontsize=14)
+ax.set_xlabel("交易日")
+ax.set_ylabel("价格 (USD)")
+ax.legend(loc="upper left")
+plt.tight_layout()
+plt.savefig("aapl_forecast.png", dpi=150)
+print()
+print("图表已保存: aapl_forecast.png", flush=True)
+print("✅ 完成！", flush=True)