AIAlpha develops a multilayer neural network architecture to predict stock returns. The project aims to understand the core principles of building such models and the nuances of training individual components. It follows a modified workflow of acquire, preprocess, train, test, and monitor, drawing inspiration from financial machine learning techniques. Bar sampling, feature engineering with moving averages and volatilities, and dimensionality reduction using autoencoders are core aspects of the approach. LSTM and Random Forest models are implemented to classify stock direction.
This project showcases a practical application of stacked autoencoders for dimensionality reduction, offering a method to handle high-dimensional datasets common in financial analysis. It demonstrates the application of both LSTM and Random Forest models for time series prediction, providing a comparative analysis of different modeling approaches. The project highlights the importance of data sampling techniques for improved model performance in financial data.
- Data Sampling: Implements bar sampling techniques based on ticks, volume, and dollar value to address issues with noise and non-stationarity in tick-level data.
- Feature Engineering: Creates features based on moving averages and rolling volatilities to enhance the predictive power of the models.
- Autoencoder: Uses a stacked autoencoder for dimensionality reduction, extracting important features from the high-dimensional dataset.
- LSTM Model: Employs an LSTM regression model for time series prediction, examining challenges related to overfitting and initialization.
- Random Forest Model: Utilizes a Random Forest classification model for classifying the direction of stock movements, providing a simpler alternative to deep learning.
- Model Evaluation: Presents results from both LSTM and Random Forest models, comparing their performance using log loss metrics.
- Educational Focus: Designed as an educational project to illustrate the application of machine learning in financial forecasting.
The project appears to be a research and educational effort. The repository includes code for data preprocessing, model building, and evaluation. The documentation provides a detailed overview of the project's methodology and implementation. However, the project involves a limited dataset size and may not be directly applicable for live trading without modifications. Recent commits indicate ongoing development and experimentation.
AIAlpha benefits those interested in applying machine learning to financial forecasting, particularly those seeking to understand stacked autoencoders and recurrent neural networks for time series prediction. It provides a practical implementation of data preprocessing, feature engineering, and model building techniques. The project is valuable for learning about the challenges and nuances of building predictive models with financial data, emphasizing the importance of data representation for improved model performance.