Update utils.py

utils.py

@@ -157,6 +157,7 @@ def format_large_number(num):
 @spaces.GPU(duration=120)
 def predict_prices(data, model=None, tokenizer=None, prediction_days=30):
     try:
+        # Panggil pipeline di sini untuk memastikan instance baru tiap run (mencegah error memori/state)
         pipeline = Chronos2Pipeline.from_pretrained("amazon/chronos-2", device_map="auto")
         
         # Chronos-2 with Covariate: Menggunakan Close (target) dan Volume (covariate)
@@ -164,16 +165,21 @@ def predict_prices(data, model=None, tokenizer=None, prediction_days=30):
         context_df.columns = ['timestamp', 'target', 'volume'] 
         context_df['id'] = 'stock_price' 
 
-        # Fix Error: Could not infer frequency (membuat data menjadi frekuensi harian)
+        # Fix Error: Could not infer frequency & FIX VOLUME COVARIATE IMPUTATION
         context_df['timestamp'] = pd.to_datetime(context_df['timestamp'])
-        context_df = context_df.set_index('timestamp').asfreq('D').fillna(method='ffill').reset_index()
+        context_df = context_df.set_index('timestamp').asfreq('D')
+        
+        # IMPUTATION FIX: Target ffill, Covariate (Volume) fillna(0)
+        context_df['target'] = context_df['target'].fillna(method='ffill')
+        context_df['volume'] = context_df['volume'].fillna(0) 
+        
+        context_df = context_df.reset_index()
         
         # Pastikan kolom sesuai urutan Chronos-2: timestamp, target, covariate(s), id
         context_df['id'] = 'stock_price'
         context_df = context_df[['timestamp', 'target', 'volume', 'id']] 
 
         with torch.no_grad():
-            # Menggunakan multiple quantile levels untuk Probabilistic Forecast (90% CI)
             pred_df = pipeline.predict_df(
                 context_df,
                 prediction_length=prediction_days,
@@ -183,12 +189,13 @@ def predict_prices(data, model=None, tokenizer=None, prediction_days=30):
                 quantile_levels=[0.1, 0.5, 0.9] 
             )
 
-        # --- FIX UTAMA: Pengecekan kolom hasil prediksi ---
+        # --- FIX UTAMA: Pengecekan kolom hasil prediksi yang lebih ketat ---
         required_cols = ['target_0.1', 'target_0.5', 'target_0.9']
         if pred_df.empty or not all(col in pred_df.columns for col in required_cols):
-             # Menggunakan Exception agar tertangkap oleh block di bawah
-             raise RuntimeError(f"Chronos-2 prediction failed or returned incomplete data. Missing columns: {[col for col in required_cols if col not in pred_df.columns]}")
-        # ----------------------------------------------------
+             # Jika gagal, pastikan kita tahu errornya dan melempar Runtime yang akan ditangkap di luar
+             missing = [col for col in required_cols if col not in pred_df.columns]
+             raise RuntimeError(f"Prediction failed. Result DataFrame is empty or incomplete. Missing: {missing}")
+        # ------------------------------------------------------------------
 
         # Ekstraksi hasil prediksi kuantil
         q05_forecast = pred_df['target_0.5'].values.astype(np.float32)
@@ -198,13 +205,11 @@ def predict_prices(data, model=None, tokenizer=None, prediction_days=30):
         
         last_price = data['Close'].iloc[-1]
         
-        # Statistik prediksi dihitung dari median (Q0.5)
         predicted_high = float(np.max(q05_forecast))
         predicted_low = float(np.min(q05_forecast))
         predicted_mean = float(np.mean(q05_forecast))
         change_pct = ((predicted_mean - last_price) / last_price) * 100 if last_price != 0 else 0
         
-        # Mengembalikan semua kuantil untuk chart
         return {
             'values': q05_forecast, 
             'dates': predicted_dates,