This project is an attempt to predict stock index prices using various deep learning techniques, specifically focusing on:
- MLP (Multilayer Perceptron)
- CNN1D (1D Convolutional Neural Networks)
- LSTM (Long Short-Term Memory Networks)
The input features and structure used for prediction include:
- Using 30 days in the past to predict 10 days in the future
- Prices
- Moving Averages
- Price Differences
The Blue vertical lines represent the start of forecasting phase
After extensive testing, it has become clear that using only prices, moving averages, and price differences as features from the dataset is insufficient for accurately predicting index prices, regardless of the deep learning method used. The inherent complexity of financial markets makes it difficult to forecast based on these limited features alone.
Among the models tested, the MLP yielded the best results in terms of minimizing the Mean Squared Error (MSE) loss during forecasting. However, even the MLP's performance was inadequate for practical use.
- Predicting index prices using these features does not provide meaningful results to real-world cases, highlighting the limitations of the dataset and feature selection.
- Despite the MLP's better performance compared to CNN1D and LSTM, it still falls short of delivering accurate predictions for stock index prices.
CSV files/: The dataset used for training and testing.* Model(s)/: Contains the differentf models.notebooks: Jupyter notebooks showcasing the experiments and results.results: Logs and performance metrics for each model.
This project serves as a learning experience in deep learning techniques applied to financial data, but the results underscore the limitations of using basic price-based features for stock index forecasting. Future work should explore more sophisticated features and models to tackle this challenging problem.
Some of the techniques to explore:
- XGBoost
- Deep Reinforcement Learning
- ARIMA (AutoRegressive Integrated Moving Average)