Update utils.py

utils.py

@@ -6,6 +6,9 @@ from datetime import datetime, timedelta
 import plotly.graph_objects as go
 from plotly.subplots import make_subplots
 import spaces
+# Impor Chronos2Pipeline untuk model amazon/chronos-2
+from chronos import Chronos2Pipeline 
+
 
 def get_indonesian_stocks():
     return {
@@ -156,26 +159,46 @@ def format_large_number(num):
 @spaces.GPU(duration=120)
 def predict_prices(data, model=None, tokenizer=None, prediction_days=30):
     try:
-        prices = data['Close'].values.astype(np.float32)
-        from chronos import BaseChronosPipeline
-        pipeline = BaseChronosPipeline.from_pretrained("amazon/chronos-bolt-base", device_map="auto")
+        # PENGGANTIAN: Gunakan Chronos2Pipeline dan model amazon/chronos-2
+        pipeline = Chronos2Pipeline.from_pretrained("amazon/chronos-2", device_map="auto")
+        
+        # Chronos-2 membutuhkan input dalam format DataFrame (predict_df)
+        context_df = data[['Close']].reset_index()
+        context_df.columns = ['timestamp', 'target'] 
+        context_df['id'] = 'stock_price' 
+
         with torch.no_grad():
-            # Fix: Use context_tensor instead of context
-            forecast = pipeline.predict(context_tensor=torch.tensor(prices), prediction_length=prediction_days)
-        forecast_np = forecast.squeeze().cpu().numpy() if isinstance(forecast, torch.Tensor) else np.array(forecast)
-        if forecast_np.ndim > 1:
-            mean_forecast = forecast_np.mean(axis=tuple(range(forecast_np.ndim - 1)))
-        else:
-            mean_forecast = forecast_np
-        last_price = prices[-1]
+            # Menggunakan predict_df()
+            pred_df = pipeline.predict_df(
+                context_df,
+                prediction_length=prediction_days,
+                id_column="id",
+                timestamp_column="timestamp",
+                target="target",
+                # Kita ambil kuantil 0.5 (median) sebagai prediksi mean
+                quantile_levels=[0.5] 
+            )
+
+        # Ekstraksi hasil prediksi dari DataFrame
+        mean_forecast = pred_df['target_0.5'].values.astype(np.float32)
+        predicted_dates = pred_df['timestamp']
+        
+        last_price = data['Close'].iloc[-1]
         predicted_high = float(np.max(mean_forecast))
         predicted_low = float(np.min(mean_forecast))
         predicted_mean = float(np.mean(mean_forecast))
         change_pct = ((predicted_mean - last_price) / last_price) * 100 if last_price != 0 else 0
-        return {'values': mean_forecast, 'dates': pd.date_range(start=data.index[-1] + timedelta(days=1), periods=len(mean_forecast), freq='D'), 'high_30d': predicted_high, 'low_30d': predicted_low, 'mean_30d': predicted_mean, 'change_pct': change_pct, 'summary': f"AI Model: Amazon Chronos-Bolt (Base)\nPredicted High: {predicted_high:.2f}\nPredicted Low: {predicted_low:.2f}\nExpected Change: {change_pct:.2f}%"}
+        
+        # Perbarui summary model
+        return {'values': mean_forecast, 'dates': predicted_dates, 'high_30d': predicted_high, 'low_30d': predicted_low, 'mean_30d': predicted_mean, 'change_pct': change_pct, 'summary': f"AI Model: Amazon Chronos-2\nPredicted High: {predicted_high:.2f}\nPredicted Low: {predicted_low:.2f}\nExpected Change: {change_pct:.2f}%"}
+        
     except Exception as e:
+        # Error handling yang lebih baik
+        error_message = f'Model error: {e}'
+        if "context_tensor" in str(e) or "context" in str(e) or "pipeline" in str(e):
+             error_message = f"Prediction API Error (Chronos-2): Cek instalasi 'chronos-forecasting' dan argumen di predict_prices(). Detail: {e}"
         print(f"Error in prediction: {e}")
-        return {'values': [], 'dates': [], 'high_30d': 0, 'low_30d': 0, 'mean_30d': 0, 'change_pct': 0, 'summary': f'Model error: {e}'}
+        return {'values': [], 'dates': [], 'high_30d': 0, 'low_30d': 0, 'mean_30d': 0, 'change_pct': 0, 'summary': error_message}
 
 def create_prediction_chart(data, predictions):
     if not len(predictions['values']):
@@ -183,12 +206,14 @@ def create_prediction_chart(data, predictions):
     fig = go.Figure()
     fig.add_trace(go.Scatter(x=data.index[-60:], y=data['Close'].values[-60:], name='Historical Price', line=dict(color='blue', width=2)))
     fig.add_trace(go.Scatter(x=predictions['dates'], y=predictions['values'], name='AI Prediction', line=dict(color='red', width=2, dash='dash')))
-    pred_std = np.std(predictions['values'])
+    # Perhitungan band harus menggunakan prediksi yang sudah dikuantil (jika ada) atau distandardisasi
+    # Karena kita hanya menggunakan satu kuantil (0.5), kita asumsikan pred_std kecil
+    pred_std = np.std(predictions['values']) if len(predictions['values']) > 1 else 0.05 * predictions['values'][0]
     upper_band = predictions['values'] + (pred_std * 1.96)
     lower_band = predictions['values'] - (pred_std * 1.96)
     fig.add_trace(go.Scatter(x=predictions['dates'], y=upper_band, name='Upper Band', line=dict(color='lightcoral', width=1)))
     fig.add_trace(go.Scatter(x=predictions['dates'], y=lower_band, name='Lower Band', line=dict(color='lightcoral', width=1), fill='tonexty', fillcolor='rgba(255,182,193,0.2)'))
-    fig.update_layout(title=f'Price Prediction - Next {len(predictions["dates"])} Days', xaxis_title='Date', yaxis_title='Price (IDR)', hovermode='x unified', height=500)
+    fig.update_layout(title=f'Price Prediction - Next {len(predictions["dates"])} Days (Chronos-2)', xaxis_title='Date', yaxis_title='Price (IDR)', hovermode='x unified', height=500)
     return fig
 
 def create_price_chart(data, indicators):