Recherche développeur

Détails du projet :

Coder un projet de programmation

Détails des prestations attendues :

SMA_50, EMA_20, OBV, Close Price (normalized), Daily Volume (normalized), Lagged Log Return (t-1), ATR: SEE MORE DETAILS ON THE NEXT PAGE Two targets: 1) Next-Day Percentage Return (Regression) and 2) Next-Day Direction (Up/Down Binary, Classification). Tool: pandas-ta OR ta 3) Preprocessing: Scale all input features using MinMaxScaler (to have uniform and proportional data influence). Convert the final DataFrame into the 3D array ([samples, time steps, features]) required by the LSTM. An interesting Array I am considering: ([1700, 64, 12]) 1764 trading days (approximately 7 years * 252 trading days/year) – 64 = 1700 days, 64 days look-back window (approximately 3 months * 21 trading days/month), 12 indicators Tool: sklearn.preprocessing.Min-MaxScaler and numpy reshaping 4) Model implementation: Build a Sequential model with 2-3 Stacked LSTM layers (by connecting multiple LSTM layers, the model’s predictive power is enhanced). Use a Dense output layer with a linear activation for the return magnitude prediction. (For directional sign (classification): If the predicted percentage return is positive, then the stock is predicted to go Up else the stock is predicted to go Down) Tool: tensorflow.keras 5) Training: Train the LSTM on the data. Training period: 01/01/2018 to 31/12/2022 (5 years) Testing period: 01/01/2023 to 31/12/2024 (2 years) 6) Evaluation: Use following models as baseline (for final comparison): Linear baseline, Random Forest AND XGBoost Compare both models using RMSE/MAE (for magnitude) and Accuracy/

Détails des exigences :

SMA_50, EMA_20, OBV, Close Price (normalized), Daily Volume (normalized), Lagged Log Return (t-1), ATR: SEE MORE DETAILS ON THE NEXT PAGE Two targets: 1) Next-Day Percentage Return (Regression) and 2) Next-Day Direction (Up/Down Binary, Classification). Tool: pandas-ta OR ta 3) Preprocessing: Scale all input features using MinMaxScaler (to have uniform and proportional data influence). Convert the final DataFrame into the 3D array ([samples, time steps, features]) required by the LSTM. An interesting Array I am considering: ([1700, 64, 12]) 1764 trading days (approximately 7 years * 252 trading days/year) – 64 = 1700 days, 64 days look-back window (approximately 3 months * 21 trading days/month), 12 indicators Tool: sklearn.preprocessing.Min-MaxScaler and numpy reshaping 4) Model implementation: Build a Sequential model with 2-3 Stacked LSTM layers (by connecting multiple LSTM layers, the model’s predictive power is enhanced). Use a Dense output layer with a linear activation for the return magnitude prediction. (For directional sign (classification): If the predicted percentage return is positive, then the stock is predicted to go Up else the stock is predicted to go Down) Tool: tensorflow.keras 5) Training: Train the LSTM on the data. Training period: 01/01/2018 to 31/12/2022 (5 years) Testing period: 01/01/2023 to 31/12/2024 (2 years) 6) Evaluation: Use following models as baseline (for final comparison): Linear baseline, Random Forest AND XGBoost Compare both models using RMSE/MAE (for magnitude) and Accuracy/