A Bayesian-based classification framework for financial time series trend prediction.

Area of Science:

• Quantitative Finance
• Machine Learning
• Deep Learning

Background:

• Financial time series analysis is crucial for understanding market dynamics.
• Machine learning and deep neural networks offer advanced techniques for analyzing complex price data patterns.
• Traditional methods often require extensive data denoising, adding complexity to preprocessing.

Purpose of the Study:

• To present a novel tri-state labeling approach for financial price data classification.
• To classify underlying price patterns into up, down, and no-action states.
• To evaluate the performance of this labeling approach using machine and deep learning models.

Main Methods:

• A tri-state labeling algorithm classifying price data into up, down, and no-action.
• Application of machine learning and deep learning models to test the labeling algorithm's performance.
• Bayesian optimization for hyperparameter tuning of the models.
• A price trend prediction module for generating trading signals.

Main Results:

• The tri-state labeling approach effectively classifies financial price patterns.
• The framework, augmented with Bayesian optimization, shows robust performance.
• The trading performance metric, average annualized Sharpe ratio, reached approximately 2.823.
• Indication of excellent cumulative returns generated by the proposed framework.

Conclusions:

• The novel tri-state labeling method simplifies data preprocessing by including a no-action state.
• The integrated framework demonstrates superior performance in financial trend prediction and trading signal generation.
• The high Sharpe ratio confirms the framework's potential for generating significant cumulative returns in financial markets.