Upload 2 files

app.py

@@ -23,7 +23,7 @@ warnings.filterwarnings('ignore')
 # 引用您組員的預測器程式
 from Bert_predict import BertPredictor
 
-# 引用新的模型預測器 (已修正特徵欄位)
+# 引用新的模型預測器
 from model_predictor import XGBoostModel
 # ========================== 引用外部模組 END ==========================
 # ========================= 新增：交易回測模組 START =========================
@@ -418,35 +418,45 @@ class TradingBacktester:
             buy_trades = trades_df[trades_df['signal'] == 1]
             sell_trades = trades_df[trades_df['signal'] == -1]
             
-            fig.add_trace(
-                go.Scatter(
-                    x=buy_trades['date'],
-                    y=[stock_data.loc[d, 'Close'] for d in buy_trades['date']],
-                    mode='markers',
-                    name='買入點',
-                    marker=dict(color='green', size=10, symbol='triangle-up')
-                ),
-                row=3, col=1
-            )
-            
-            fig.add_trace(
-                go.Scatter(
-                    x=sell_trades['date'],
-                    y=[stock_data.loc[d, 'Close'] for d in sell_trades['date']],
-                    mode='markers',
-                    name='賣出點',
-                    marker=dict(color='red', size=10, symbol='triangle-down')
-                ),
-                row=3, col=1
-            )
+            if not buy_trades.empty:
+                fig.add_trace(
+                    go.Scatter(
+                        x=buy_trades['date'],
+                        y=buy_trades['price'],
+                        mode='markers',
+                        name='買入',
+                        marker=dict(color='red', size=8, symbol='triangle-up')
+                    ),
+                    row=3, col=1
+                )
+                
+            if not sell_trades.empty:
+                fig.add_trace(
+                    go.Scatter(
+                        x=sell_trades['date'],
+                        y=sell_trades['price'],
+                        mode='markers',
+                        name='賣出',
+                        marker=dict(color='green', size=8, symbol='triangle-down')
+                    ),
+                    row=3, col=1
+                )
         
+        # 更新布局
         fig.update_layout(
-            title='交易回測結果',
+            title=f"交易策略回測結果",
             height=800,
-            showlegend=True
+            showlegend=True,
+            xaxis3_title="日期"
         )
+        
+        fig.update_yaxes(title_text="價值 (TWD)", row=1, col=1)
+        fig.update_yaxes(title_text="股數", row=2, col=1)
+        fig.update_yaxes(title_text="股價 (TWD)", row=3, col=1)
+        
         return fig
 
+
 def create_backtest_summary_card(results):
     """創建回測摘要卡片"""
     if not results:
@@ -797,98 +807,47 @@ def simple_statistical_predict(data, predict_days=5):
 
 def calculate_new_features(df):
     """
-    計算新的技術指標特徵 - 完整對應訓練模型的特徵
+    計算新的技術指標特徵 - 針對新特徵需求
     """
     if df.empty:
         return df
     
-    # 確保基礎欄位存在
-    if 'Close' not in df.columns and 'close' in df.columns:
-        df['Close'] = df['close']
-    if 'Volume' not in df.columns and 'volume' in df.columns:
-        df['Volume'] = df['volume']
-    if 'High' not in df.columns and 'high' in df.columns:
-        df['High'] = df['high']
-    if 'Low' not in df.columns and 'low' in df.columns:
-        df['Low'] = df['low']
-    if 'Open' not in df.columns and 'open' in df.columns:
-        df['Open'] = df['open']
-    
-    # 1. close - 收盤價
-    df['close'] = df['Close']
-    
-    # 2. return_t-1 — 前一日報酬率 (***FIXED: Corrected to use hyphen to match the model***)
+    # 1. return_t-1 – 前一日報酬率
     df['return_t-1'] = df['Close'].pct_change()
     
-    # 3. return_t-5 — 過去 5 日累積報酬率 (***FIXED: Corrected to use hyphen to match the model***)
+    # 2. return_t-5 – 過去 5 日累積報酬率
     df['return_t-5'] = (df['Close'] / df['Close'].shift(5) - 1)
     
-    # 4. MA5_close — 5 日移動平均價
+    # 3. MA5_close – 5 日移動平均價
     df['MA5_close'] = df['Close'].rolling(window=5).mean()
     
-    # 5. volatility_5d — 5 日報酬標準差（短期波動）
+    # 4. MA20_close – 20 日移動平均價  
+    df['MA20_close'] = df['Close'].rolling(window=20).mean()
+    
+    # 5. volatility_5d – 5 日報酬標準差（短期波動）
     df['volatility_5d'] = df['return_t-1'].rolling(window=5).std()
     
-    # 6. volume_ratio_5d — 今日成交量 ÷ 5 日均量
+    # 6. volume_ratio_5d – 今日成交量 ÷ 5 日均量
     df['volume_5d_avg'] = df['Volume'].rolling(window=5).mean()
     df['volume_ratio_5d'] = df['Volume'] / df['volume_5d_avg']
     
-    # 7. MACD_diff — MACD - signal（趨勢強弱）
-    exp1 = df['Close'].ewm(span=12).mean()
-    exp2 = df['Close'].ewm(span=26).mean()
-    macd_line = exp1 - exp2
-    signal_line = macd_line.ewm(span=9).mean()
-    df['MACD_diff'] = macd_line - signal_line
-    
-    # 8. dji_return_t-1 — 前一日道瓊指數報酬率（預設為0，需外部數據）
-    df['dji_return_t-1'] = 0.0
-    
-    # 9. sox_return_t-1 — 前一日費半指數報酬率（預設為0，需外部數據）
-    df['sox_return_t-1'] = 0.0
-    
-    # 10. NEWS — 新聞情緒分數（預設為0，需外部數據）
-    df['NEWS'] = 0.0
-    
-    # 11. MACDvol — MACD柱狀圖
-    df['MACDvol'] = 0.0
-    
-    # 12. RSI_14 — 14日RSI
+    # 7. RSI_14 – 14 日 RSI 指標
     delta = df['Close'].diff()
     gain = (delta.where(delta > 0, 0)).rolling(window=14).mean()
     loss = (-delta.where(delta < 0, 0)).rolling(window=14).mean()
     rs = gain / loss
     df['RSI_14'] = 100 - (100 / (1 + rs))
     
-    # 13. ADX — 平均趨向指標
-    try:
-        df['up_move'] = df['High'] - df['High'].shift(1)
-        df['down_move'] = df['Low'].shift(1) - df['Low']
-        df['+DM'] = np.where((df['up_move'] > df['down_move']) & (df['up_move'] > 0), df['up_move'], 0)
-        df['-DM'] = np.where((df['down_move'] > df['up_move']) & (df['down_move'] > 0), df['down_move'], 0)
-        
-        high_low = df['High'] - df['Low']
-        high_close_prev = np.abs(df['High'] - df['Close'].shift(1))
-        low_close_prev = np.abs(df['Low'] - df['Close'].shift(1))
-        df['TR'] = np.maximum.reduce([high_low, high_close_prev, low_close_prev])
-        
-        df['+DI'] = (df['+DM'].ewm(com=13, adjust=False).mean() / df['TR'].ewm(com=13, adjust=False).mean()) * 100
-        df['-DI'] = (df['-DM'].ewm(com=13, adjust=False).mean() / df['TR'].ewm(com=13, adjust=False).mean()) * 100
-        df['DX'] = np.abs(df['+DI'] - df['-DI']) / (df['+DI'] + df['-DI']) * 100
-        df['ADX'] = df['DX'].ewm(com=13, adjust=False).mean()
-    except Exception:
-        print("警告：ADX計算失敗，使用預設值")
-        df['ADX'] = 25.0
-    
-    # 14. volume_weighted_return — 成交量加權報酬率
-    df['volume_weighted_return'] = np.abs(df['return_t-1']) * df['Volume']
+    # 8. MACD_diff – MACD - signal（趨勢強弱）
+    exp1 = df['Close'].ewm(span=12).mean()
+    exp2 = df['Close'].ewm(span=26).mean()
+    macd_line = exp1 - exp2
+    signal_line = macd_line.ewm(span=9).mean()
+    df['MACD_diff'] = macd_line - signal_line
     
     # 移除輔助欄位
-    auxiliary_columns = ['volume_5d_avg', 'up_move', 'down_move', '+DM', '-DM', 'TR', '+DI', '-DI', 'DX']
-    df.drop(columns=[col for col in auxiliary_columns if col in df.columns], inplace=True)
-    
-    # 填補 NaN 值
-    df.fillna(method='ffill', inplace=True)
-    df.fillna(0, inplace=True)
+    if 'volume_5d_avg' in df.columns:
+        df = df.drop('volume_5d_avg', axis=1)
     
     return df
 
@@ -904,92 +863,171 @@ def advanced_xgboost_predict(predict_days=5):
         xgb_model = XGBoostModel()
         
         # 獲取台指期數據 (作為主要標的)
-        taiex_data_raw = get_stock_data('^TWII', '2y')
-        if taiex_data_raw.empty or len(taiex_data_raw) < 60:
+        taiex_data = get_stock_data('^TWII', '2y')
+        if taiex_data.empty or len(taiex_data) < 60:
             print("台指期數據不足，無法進行XGBoost預測")
             return None
+        
+        # 計算技術指標（包含舊的指標）
+        taiex_data = calculate_technical_indicators(taiex_data)
+        
+        # 計算新特徵
+        taiex_data = calculate_new_features(taiex_data)
+        
+        # 獲取美股指數數據來計算外部指標
+        us_market_data = get_us_market_data()
+        
+        # 獲取新聞情緒分數
+        try:
+            if predictor is not None:
+                sentiment_score_raw = predictor.get_news_index()
+                if sentiment_score_raw is None:
+                    sentiment_score_raw = 0
+            else:
+                sentiment_score_raw = 0
+        except:
+            sentiment_score_raw = 0
+        
+        # 準備特徵數據 (使用最新的數據點)
+        latest_data = taiex_data.iloc[-1]
+        
+        # 取得昨日收盤價
+        yesterday_close = latest_data['Close']
 
-        # ***FIX START: Simplified and Corrected Feature Engineering and Assembly***
-
-        # 1. Calculate all technical features
-        taiex_data = calculate_new_features(taiex_data_raw)
-
-        # 2. Get the latest row of data which contains most of our features
-        latest_features = taiex_data.iloc[-1:].copy()
-
-        # 3. Update external features (US markets, News)
-        # Get US market returns
+        # 特徵列表，確保與模型訓練時完全一致
+        new_feature_columns = [
+            'return_t-1',
+            'return_t-5',
+            'MA5_close',
+            'volatility_5d',
+            'volume_ratio_5d',
+            'MACD_diff',
+        ]
+        
+        # 添加美股指標（如果有數據的話）
         dji_return = 0
         sox_return = 0
+        
+        # 嘗試獲取美股前一日報酬率
         try:
             dji_data = get_stock_data('^DJI', '5d')
             if not dji_data.empty and len(dji_data) >= 2:
                 dji_return = (dji_data['Close'].iloc[-1] / dji_data['Close'].iloc[-2] - 1)
-        except Exception as e:
-            print(f"Could not fetch DJI data: {e}")
+        except:
+            pass
             
         try:
             sox_data = get_stock_data('^SOX', '5d') 
             if not sox_data.empty and len(sox_data) >= 2:
                 sox_return = (sox_data['Close'].iloc[-1] / sox_data['Close'].iloc[-2] - 1)
-        except Exception as e:
-            print(f"Could not fetch SOX data: {e}")
+        except:
+            pass
+        
+        # 檢查並處理 NaN 值，建立特徵狀態記錄
+        feature_status = {}
+        features_list = []
+        feature_names = []
+        
+        # 處理本地計算的技術指標特徵
+        for feature in new_feature_columns:
+            if feature in latest_data.index:
+                value = latest_data[feature]
+                if pd.isna(value):
+                    # 使用合理的預設值
+                    if 'return' in feature: default_value = 0.0
+                    elif 'MA' in feature: default_value = latest_data['Close'] if not pd.isna(latest_data['Close']) else 100
+                    elif 'volatility' in feature: default_value = 0.02
+                    elif 'volume_ratio' in feature: default_value = 1.0
+                    elif 'MACD' in feature: default_value = 0.0
+                    else: default_value = 0.0
+                    
+                    features_list.append(default_value)
+                    feature_status[feature] = {'value': default_value, 'is_real': False, 'source': 'default'}
+                else:
+                    features_list.append(value)
+                    feature_status[feature] = {'value': value, 'is_real': True, 'source': 'calculated'}
+                
+                feature_names.append(feature)
+
+        # 按照模型訓練的順序添加剩餘特徵
+        # 7. dji_return_t-1
+        features_list.append(dji_return)
+        feature_names.append('dji_return_t-1')
+        feature_status['dji_return_t-1'] = {
+            'value': dji_return, 
+            'is_real': dji_return != 0, 
+            'source': 'calculated' if dji_return != 0 else 'default'
+        }
 
-        # Get news sentiment score
-        sentiment_score_raw = 0
-        try:
-            if predictor is not None:
-                score = predictor.get_news_index()
-                if score is not None:
-                    sentiment_score_raw = score
-        except Exception as e:
-            print(f"Could not get news score: {e}")
-
-        # Update the values in our feature set
-        latest_features['dji_return_t-1'] = dji_return
-        latest_features['sox_return_t-1'] = sox_return
-        latest_features['NEWS'] = sentiment_score_raw
-
-        # 4. Define the exact feature list from the training script
-        model_feature_columns = [
-            'close',
-            'return_t-1',
-            'return_t-5',
-            'MA5_close',
-            'volatility_5d',
-            'volume_ratio_5d',
-            'MACD_diff',
-            'dji_return_t-1',
-            'sox_return_t-1',
-            'NEWS',
-            'MACDvol',
-            'RSI_14',
-            'ADX',
-            'volume_weighted_return'
-        ]
+        # 8. sox_return_t-1
+        features_list.append(sox_return)
+        feature_names.append('sox_return_t-1')
+        feature_status['sox_return_t-1'] = {
+            'value': sox_return, 
+            'is_real': sox_return != 0, 
+            'source': 'calculated' if sox_return != 0 else 'default'
+        }
 
-        # 5. Create the final input DataFrame with columns in the correct order
-        # This replaces the entire flawed feature assembly loop from the original code.
-        input_df = latest_features[model_feature_columns]
+        # 9. close
+        if not pd.isna(yesterday_close):
+            features_list.append(yesterday_close)
+            feature_status['close'] = {'value': yesterday_close, 'is_real': True, 'source': 'calculated'}
+        else:
+            features_list.append(10000) # Fallback value for price
+            feature_status['close'] = {'value': 10000, 'is_real': False, 'source': 'default'}
+        feature_names.append('close')
 
-        # Log the final features for verification
+        # 10. NEWS
+        features_list.append(sentiment_score_raw)
+        feature_status['NEWS'] = {'value': sentiment_score_raw, 'is_real': True, 'source': 'calculated'}
+        feature_names.append('NEWS')
+        
+        # 轉換為 DataFrame (XGBoost 模型期望的格式)
+        input_df = pd.DataFrame([features_list], columns=feature_names)
+        
+        # 詳細的資料驗證日誌
         print("=" * 60)
-        print("XGBoost Model Input Features Verification")
+        print("XGBoost 模型輸入特徵檢查報告 (漲幅百分比版��)")
         print("=" * 60)
-        print(input_df.iloc[0])
+        
+        print(f"總特徵數量: {len(features_list)} 個")
+        print(f"新聞情緒分數: {sentiment_score_raw:.6f}")
+        
+        # 特徵詳細狀態
+        print("\n特徵狀態詳情:")
+        for i, (name, value) in enumerate(zip(feature_names, features_list)):
+            status = feature_status.get(name, {})
+            status_symbol = "✓正常" if status.get('is_real', False) else "⚠ 預設值"
+            print(f"  [{i+1:2d}] {name:18s}: {value:12.6f} ({status_symbol})")
+        
+        # 統計完整性
+        real_features = sum(1 for status in feature_status.values() if status.get('is_real', False))
+        total_features = len(feature_status)
+        completeness = (real_features / total_features) * 100 if total_features > 0 else 0
+        
+        print(f"\n特徵完整性:")
+        print(f"  實際計算特徵: {real_features}/{total_features} ({completeness:.1f}%)")
+        if completeness < 70:
+            print("  警告: 超過30%的特徵使用預設值，可能影響預測準確性")
+        else:
+            print("  特徵完整性良好")
+        
+        # 顯示完整特徵向量
+        print(f"\n完整特徵向量 (共{len(features_list)}個特徵):")
+        for i, (name, value) in enumerate(zip(feature_names, features_list)):
+            print(f"  [{i+1:2d}] {name:18s}: {value:12.6f}")
+        
         print("=" * 60)
         
-        # ***FIX END***
-
         # 進行預測
         predictions = xgb_model.predict('xgboost_model', input_df)
         
-        # 處理新的漲幅百分比輸出格式
+        # 【重要更新】處理新的漲幅百分比輸出格式
         pred_mapping = {
-            1: 'Change_pct_t1_pred',
-            5: 'Change_pct_t5_pred',
-            10: 'Change_pct_t10_pred',
-            20: 'Change_pct_t20_pred'
+            1: 'Change_pct_t1_pred',   # 1天後漲幅%
+            5: 'Change_pct_t5_pred',   # 5天後漲幅%
+            10: 'Change_pct_t10_pred', # 10天後漲幅%
+            20: 'Change_pct_t20_pred'  # 20天後漲幅%
         }
         
         # 找到最接近的預測天數
@@ -997,24 +1035,25 @@ def advanced_xgboost_predict(predict_days=5):
         closest_day = min(available_days, key=lambda x: abs(x - predict_days))
         pred_key = pred_mapping[closest_day]
         
+        # 【關鍵修改】現在直接取得漲幅百分比
         predicted_change_pct = predictions[pred_key]
         
-        current_price = taiex_data['Close'].iloc[-1]
+        # 【新增】為了兼容性，計算預測價格（僅供參考）
+        current_price = latest_data['Close']
         predicted_price = current_price * (1 + predicted_change_pct / 100)
         
-        # Use a simple confidence score for now
-        confidence = 0.8 
-
         print(f"XGBoost 預測完成:")
         print(f"- 預測天數: {predict_days} (使用 {closest_day} 天模型)")
         print(f"- 當前價格: {current_price:.2f}")
         print(f"- 預測漲幅: {predicted_change_pct:+.2f}%")
         print(f"- 預測價格: {predicted_price:.2f} (參考)")
+        print(f"- 使用特徵數: {len(features_list)} 個")
+        print(f"- 特徵完整性: {completeness:.1f}%")
         
         return {
-            'predicted_price': predicted_price,
-            'change_pct': predicted_change_pct,
-            'confidence': confidence
+            'predicted_price': predicted_price,     # 為了兼容現有代碼
+            'change_pct': predicted_change_pct,     # 【新增】直接的漲幅百分比
+            'confidence': max(0.6, min(0.85, completeness / 100))  # 根據特徵完整性調整信心度
         }
         
     except Exception as e:

model_predictor.py

@@ -30,11 +30,7 @@ class XGBoostModel:
             'MACD_diff',         # MACD - signal
             'dji_return_t-1',    # 前一日道瓊指數報酬率
             'sox_return_t-1',    # 前一日費半指數報酬率
-            'NEWS',              # 新聞情緒分數
-            'MACDvol',           # MACD 柱狀圖
-            'RSI_14',            # 14 日 RSI
-            'ADX',               # 平均趨向指標
-            'volume_weighted_return'  # 成交量加權報酬率
+            'NEWS'               # 新聞情緒分數
         ]
         
         # 【新增】輸出目標對應表
@@ -103,98 +99,6 @@ class XGBoostModel:
             self.scaler = StandardScaler()
             return False
 
-    def create_features_from_stock_data(self, stock_data):
-        """
-        從股票資料創建所需的特徵
-        完全對應訓練腳本中的 create_new_features 函數
-        
-        Args:
-            stock_data: yfinance 格式的股票資料 DataFrame
-            
-        Returns:
-            processed_df: 包含所有特徵的 DataFrame
-        """
-        df = stock_data.copy()
-        
-        # 確保必要的基礎欄位存在
-        required_base_columns = ['Close', 'Volume', 'High', 'Low']
-        for col in required_base_columns:
-            if col not in df.columns:
-                raise ValueError(f"缺少必要的基礎欄位: {col}")
-        
-        # 統一欄位名稱（yfinance 使用大寫）
-        df['close'] = df['Close']
-        df['volume'] = df['Volume']
-        
-        # 1. return_t-1 — 前一日報酬率
-        df['return_t-1'] = df['close'].pct_change()
-        
-        # 2. return_t-5 — 過去 5 日累積報酬率
-        df['return_t-5'] = (df['close'] / df['close'].shift(5) - 1)
-        
-        # 3. MA5_close — 5 日移動平均價
-        df['MA5_close'] = df['close'].rolling(window=5).mean()
-        
-        # 4. volatility_5d — 5 日報酬標準差
-        df['volatility_5d'] = df['return_t-1'].rolling(window=5).std()
-        
-        # 5. volume_ratio_5d — 今日成交量 ÷ 5 日均量
-        df['volume_5d_avg'] = df['volume'].rolling(window=5).mean()
-        df['volume_ratio_5d'] = df['volume'] / df['volume_5d_avg']
-        
-        # 6. MACD_diff — MACD - signal
-        exp1 = df['close'].ewm(span=12).mean()
-        exp2 = df['close'].ewm(span=26).mean()
-        macd_line = exp1 - exp2
-        signal_line = macd_line.ewm(span=9).mean()
-        df['MACD_diff'] = macd_line - signal_line
-        
-        # 7-8. 美股指數報酬率（需要外部資料，暫設為0）
-        df['dji_return_t-1'] = 0.0
-        df['sox_return_t-1'] = 0.0
-        
-        # 9. NEWS — 新聞情緒分數（需要外部資料，暫設為0）
-        df['NEWS'] = 0.0
-        
-        # 10. MACDvol — MACD柱狀圖（需要外部資料，暫設為0）
-        df['MACDvol'] = 0.0
-        
-        # 11. RSI_14 — 14日RSI
-        delta = df['close'].diff()
-        gain = (delta.where(delta > 0, 0)).rolling(window=14).mean()
-        loss = (-delta.where(delta < 0, 0)).rolling(window=14).mean()
-        rs = gain / loss
-        df['RSI_14'] = 100 - (100 / (1 + rs))
-        
-        # 12. ADX — 平均趨向指標
-        df['up_move'] = df['High'] - df['High'].shift(1)
-        df['down_move'] = df['Low'].shift(1) - df['Low']
-        df['+DM'] = np.where((df['up_move'] > df['down_move']) & (df['up_move'] > 0), df['up_move'], 0)
-        df['-DM'] = np.where((df['down_move'] > df['up_move']) & (df['down_move'] > 0), df['down_move'], 0)
-        
-        high_low = df['High'] - df['Low']
-        high_close_prev = np.abs(df['High'] - df['close'].shift(1))
-        low_close_prev = np.abs(df['Low'] - df['close'].shift(1))
-        df['TR'] = np.maximum.reduce([high_low, high_close_prev, low_close_prev])
-        
-        df['+DI'] = (df['+DM'].ewm(com=13, adjust=False).mean() / df['TR'].ewm(com=13, adjust=False).mean()) * 100
-        df['-DI'] = (df['-DM'].ewm(com=13, adjust=False).mean() / df['TR'].ewm(com=13, adjust=False).mean()) * 100
-        df['DX'] = np.abs(df['+DI'] - df['-DI']) / (df['+DI'] + df['-DI']) * 100
-        df['ADX'] = df['DX'].ewm(com=13, adjust=False).mean()
-        
-        # 13. volume_weighted_return — 成交量加權報酬率
-        df['volume_weighted_return'] = np.abs(df['return_t-1']) * df['volume']
-        
-        # 清理輔助欄位
-        cleanup_columns = ['volume_5d_avg', 'up_move', 'down_move', '+DM', '-DM', 'TR', '+DI', '-DI', 'DX']
-        df.drop(columns=[col for col in cleanup_columns if col in df.columns], inplace=True)
-        
-        # 填補 NaN 值
-        df.fillna(method='ffill', inplace=True)
-        df.fillna(0, inplace=True)  # 剩餘的 NaN 用 0 填補
-        
-        return df
-    
     def preprocess_features(self, input_df):
         """
         預處理輸入特徵
@@ -496,11 +400,7 @@ if __name__ == "__main__":
         'MACD_diff': [0.5],
         'dji_return_t-1': [0.01],
         'sox_return_t-1': [0.015],
-        'NEWS': [0.1],
-        'MACDvol': [0.2],
-        'RSI_14': [55.0],
-        'ADX': [25.0],
-        'volume_weighted_return': [1000.0]
+        'NEWS': [0.1]
     })
     
     print("測試模型預測器...")