將prepare_ml_data的df = df.copy()移除

def prepare_ml_data(df, return_threshold=0.003):
    df['Future_Close'] = df['close'].shift(-1)
    df['Return'] = (df['Future_Close'] / df['close']) - 1
    df['Target'] = (df['Return'] > return_threshold).astype(int)

    features = [
        'MA20', 'MA50', 'RSI', 'MACD', 'Signal',
        'BB_Mid', 'BB_Upper', 'BB_Lower',
        'Vol_MA20', 'Vol_MA50',
        'ROC', 'ATR', 'Momentum',
        'Price_vs_MA20', 'Price_vs_MA50'
    ]
    df = df.dropna().reset_index(drop=True)
    X = df[features]
    y = df['Target']
    return X, y, df

優勢： 提高代碼執行效率，避免不必要的數據複製。

在 walk_forward_train 函式中，當 folds 列表不為空時，加入一個檢查，確保 np.mean(scores) 不會因為 scores 為空而報錯

# -------------------------
# Step 4: 滑動視窗 Walk-forward 訓練 (修正版)
# -------------------------
def walk_forward_train(X, y, df, train_window=2000, test_window=400, step=400, model_cls=RandomForestClassifier, model_kwargs=None):
    """
    sliding-window walk-forward training.
    """
    if model_kwargs is None:
        model_kwargs = {"n_estimators":200, "random_state":42}

    n = len(X)
    folds = []
    scores = []
    last_model = None
    last_test_index = None
    last_y_true = None
    last_y_pred = None

    start = 0
    # 從 train_window 開始，確保有足夠訓練資料
    while start + train_window + test_window <= n:
        train_idx = list(range(start, start + train_window))
        test_idx = list(range(start + train_window, start + train_window + test_window))

        X_train, X_test = X.iloc[train_idx], X.iloc[test_idx]
        y_train, y_test = y.iloc[train_idx], y.iloc[test_idx]

        model = model_cls(**model_kwargs)
        model.fit(X_train, y_train)
        y_pred = model.predict(X_test)
        acc = accuracy_score(y_test, y_pred)
        scores.append(acc)
        
        # 為了簡化輸出，這裡不列印每個 Fold 的細節
        # print(f"Fold {len(scores)} Accuracy: {acc:.4f}")

        folds.append({
            "train_index": train_idx,
            "test_index": test_idx,
            "model": model,
            "y_true": y_test,
            "y_pred": y_pred,
            "acc": acc
        })

        # move window
        start += step

        # 保留最後一個 fold 的結果
        last_model = model
        last_test_index = test_idx
        last_y_true = y_test
        last_y_pred = y_pred

    print(f"\nWalk-forward 訓練完成。共 {len(scores)} 個 Fold。")

    # 📌 修正點：在計算平均準確率之前檢查 scores 列表是否為空
    if scores:
        avg_acc = np.mean(scores)
        print(f"📊 平均準確率: {avg_acc:.4f}")
    else:
        # 如果 scores 為空，則平均準確率為 N/A
        print("📊 平均準確率: N/A (無 Fold 執行)")

    if not folds:
        print("❌ 錯誤：Walk-forward 沒有執行任何 Fold。請檢查資料量是否足夠。")
        return None, None, None, None, None
   
    # 🎯 新增分類報告，以量化模型在類別不平衡下的性能
    from sklearn.metrics import classification_report # 確保這裡也可以使用

    if last_y_pred is not None and len(last_y_pred) > 0:
        print("\n=== 最後一個 Walk-forward 區段的詳細分類報告 (Target=1: 預測漲幅>0.3%) ===")
        print(classification_report(last_y_true, last_y_pred, 
                                     target_names=['Target=0 (Down/Small)', 'Target=1 (Up)'], 
                                     zero_division=0))

    # 使用最後一個 Fold 的結果繪製預測圖
    plot_predictions(df, last_y_true, last_y_pred, last_test_index, f"Predictions vs Actual (Latest {test_window} bars)")

    return last_model, last_test_index, last_y_true, last_y_pred, folds

優勢： 雖然在 if not folds: 已經處理了，但更好的做法是確保 np.mean 的輸入始終有效，提高代碼的魯棒性。