Try to fix the AI some more

WebServer/AIPython/aipredictor.py

@@ -49,20 +49,29 @@ def Predict():
     scaled_predictions = reconstructed_model.predict(scaled_data)
 
     # Use the loaded target scaler to get back to % change
-    actual_prediction = target_scaler.inverse_transform(scaled_predictions)
+    actual_prediction = target_scaler.inverse_transform(scaled_predictions.reshape(-1, 1)).flatten()
 
     # 'predictions' will be a 2D array, flatten it if you want a simple list
-    flat_predictions = actual_prediction.flatten().tolist()
+    flat_predictions = actual_prediction
+
+    # Get the overall trend to pull predictions from
+    predictionTrend = 0
+    with open("Target_Close_Average.txt", "r") as f:
+        predictionTrend = float(f.read().strip())
 
     # Set the movement indicator
     movement_indicator = 0
-    if (np.mean(flat_predictions) > 0.01):
+    averagePrediction = np.mean(flat_predictions)
+    if (averagePrediction > 0.005): # as in 3% swing up
         movement_indicator = 1
-    elif (np.mean(flat_predictions) < -0.01):
+    elif (averagePrediction < -0.005): # as in 3% swing down
         movement_indicator = -1
     else:
         movement_indicator = 0
 
+    # Debug data
+    print(f"averagePrediction: {averagePrediction}")
+
     # Return to C# via stdout
     print(f"---RESULT_START---")
     print(movement_indicator)

WebServer/AIPython/aitrainer.py

@@ -35,12 +35,14 @@ def TrainAI():
     # Create the DNN
     dnn_model = Sequential([
         layers.Input(shape=(train_features.shape[1],)), # Load the feature count dynamically
-        layers.Dense(256, activation='elu'),
+        layers.Dense(256, activation='elu'),            # DNN layer
-        layers.Dropout(0.1),                    # Small dropout to prevent it from memorizing noise
+        layers.BatchNormalization(),                    # Nomralize
-        layers.Dense(128, activation='elu'),
+        layers.Dropout(0.1),                            # Small dropout to prevent it from memorizing noise
-        layers.Dropout(0.1),                    # Small dropout to prevent it from memorizing noise
+        layers.Dense(128, activation='elu'),            # DNN layer
-        layers.Dense(24, activation='elu'),
+        layers.BatchNormalization(),                    # Nomralize
-        layers.Dense(1, activation='linear')
+        layers.Dropout(0.1),                            # Small dropout to prevent it from memorizing noise
+        layers.Dense(24, activation='elu'),             # DNN layer
+        layers.Dense(1, activation='linear')            # DNN layer
     ])
 
     # Allow negative numbers

WebServer/AIPython/features.py

@@ -11,37 +11,25 @@ def MakeFeatures(df):
     df['Body_Size'] = (df['Close'] - df['Open']).abs() / (df['High'] - df['Low'])
     df['Upper_Shadow'] = (df['High'] - candle_top) / (df['High'] - df['Low'])
 
-# Remove Unused Symbols to save ram
-    df.drop('Open', axis=1, inplace=True)
-    df.drop('High', axis=1, inplace=True)
-    df.drop('Low', axis=1, inplace=True)
-
      # Is volume 2x higher than the 20-day average?
     df['Vol_Intensity'] = df['Volume'] / df['Volume'].shift(1).rolling(20).mean()
     # Volume Change
     df['Volume_Chg'] = df['Volume'].pct_change()
 
-# Remove Unused Symbols to save ram
-    df.drop('Volume', axis=1, inplace=True)
-
     # Moving Average Crossover (Golden/Death Cross logic)
-    df['Moving_Average_5'] = df['Close'].shift(1).rolling(window=20).mean()
+    df['Moving_Average_5'] = df['Close'].rolling(window=5).mean()
-    df['Moving_Average_20'] = df['Close'].shift(1).rolling(window=20).mean()
+    df['Moving_Average_20'] = df['Close'].rolling(window=20).mean()
     # if short term > long term (bullish), else 0
     df['Trend_Signal'] = (df['Moving_Average_5'] > df['Moving_Average_20']).astype(int)
     # Distance from MA (How overextended are we?)
     df['Dist_From_MA20'] = (df['Close'] / df['Moving_Average_20']) - 1
 
     # Bollinger Band Position (Where are we relative to volatility?)
-    std_20 = df['Close'].shift(1).rolling(20).std()
+    std_20 = df['Close'].rolling(20).std()
     upper_band = df['Moving_Average_20'] + (std_20 * 2)
     lower_band = df['Moving_Average_20'] - (std_20 * 2)
     df['BB_Pos'] = (df['Close'] - lower_band) / (upper_band - lower_band)
 
-# Remove Unused Symbols to save ram
-    df.drop('Moving_Average_5', axis=1, inplace=True)
-    df.drop('Moving_Average_20', axis=1, inplace=True)
-
     # Add feature for Returns
     df['Return'] = df['Close'].pct_change()
     # Log Returns (Better for AI than pct_change for statistical normality)
@@ -65,10 +53,17 @@ def MakeFeatures(df):
         df[f'Vol_Lag_{lag}'] = df['Volume_Chg'].shift(lag)
 
     # This is our training metric of price difference 5 days ahead
-    df['Target_Close'] = np.log(df['Close'].shift(-5) / df['Close']) / df['Volatility_5']
+    df['Target_Close'] = (np.log(df['Close'].shift(-5) / df['Close']) / df['Volatility_5']).clip(-10, 10)
 
-# Remove Unused Symbols to save ram
+    # Save the overall trend to predict based off of later
-    df.drop('Close', axis=1, inplace=True)
+    with open("Target_Close_Average.txt", "w") as file:
+        file.write(str(df["Target_Close"].mean()))
+
+    # Drop every column that is a raw price or an unscaled average
+    cols_to_drop = ['Open', 'High', 'Low', 'Volume', 'Close', 'Moving_Average_5', 'Moving_Average_20']
+    for col in cols_to_drop:
+        if col in df.columns:
+            df.drop(col, axis=1, inplace=True)
 
     # Drop rows with null values
     df.dropna(inplace=True)