Forecasting Financial Time Series through Causal and Dilated Convolutional Neural Networks

Area of Science:

• Quantitative Finance
• Machine Learning Applications
• Time Series Analysis

Background:

• Accurate financial market prediction remains challenging due to market complexity and temporal data dependencies.
• Traditional models often rely on efficient market hypothesis assumptions, which may not fully capture market dynamics.
• The temporal structure of financial data significantly impacts model performance, requiring careful data handling.

Purpose of the Study:

• To adapt a novel machine learning model, successful in audio and speech generation, for financial time series prediction.
• To compare the performance of the adapted model against a naive base model constrained by efficient market assumptions.
• To evaluate the model's efficacy in predicting future stock index price movements and trends.

Main Methods:

• A generative model, previously used for audio and speech, was modified for financial data analysis.
• The model was trained and validated using historical stock index data, considering temporal data splits.
• Performance was benchmarked against a naive model adhering to efficient market hypothesis principles.

Main Results:

• The adapted model demonstrated superior predictive performance compared to the naive base model.
• The model achieved over 20% improvement in predicting the next day's closing price.
• The model showed a 37% improvement in predicting the next day's price trend.

Conclusions:

• The adapted generative model shows significant potential for improving financial market prediction accuracy.
• The study highlights the importance of handling temporal data structures in financial time series analysis.
• This approach offers a promising alternative to traditional models for forecasting stock index movements.