[Paper Review] Supervised Autoencoder MLP for Financial Time Series Forecasting

Bartosz Bieganowski, Robert Ślepaczuk|arXiv (Cornell University)|Apr 2, 2024

Stock Market Forecasting MethodsDecision Sciences3 citations

TL;DR

This paper proposes a Supervised Autoencoder Multi-Layer Perceptron (SAE-MLP) for financial time series forecasting, using noise-augmented data and triple barrier labeling to improve risk-adjusted returns. It demonstrates that optimal noise levels and bottleneck sizes significantly boost strategy performance, with hyperparameter tuning proving critical for outperforming traditional approaches on S&P 500, EUR/USD, and BTC/USD data.

ABSTRACT

This paper investigates the enhancement of financial time series forecasting with the use of neural networks through supervised autoencoders, aiming to improve investment strategy performance. It specifically examines the impact of noise augmentation and triple barrier labeling on risk-adjusted returns, using the Sharpe and Information Ratios. The study focuses on the S&P 500 index, EUR/USD, and BTC/USD as the traded assets from January 1, 2010, to April 30, 2022. Findings indicate that supervised autoencoders, with balanced noise augmentation and bottleneck size, significantly boost strategy effectiveness. However, excessive noise and large bottleneck sizes can impair performance, highlighting the importance of precise parameter tuning. This paper also presents a derivation of a novel optimization metric that can be used with triple barrier labeling. The results of this study have substantial policy implications, suggesting that financial institutions and regulators could leverage techniques presented to enhance market stability and investor protection, while also encouraging more informed and strategic investment approaches in various financial sectors.

Motivation & Objective

• To improve algorithmic investment strategy performance using supervised autoencoders in financial time series forecasting.
• To investigate whether data augmentation via noise and denoising through autoencoders enhances strategy effectiveness.
• To evaluate the impact of triple barrier labeling versus simple direction classification on classifier performance.
• To assess the role of hyperparameter tuning in optimizing risk-adjusted returns.
• To develop a novel optimization metric tailored for triple barrier labeling in financial forecasting.

Proposed method

• Employed a Supervised Autoencoder-MLP (SAE-MLP) architecture to learn compressed representations of high-frequency price data.
• Applied Gaussian noise augmentation during training to improve generalization and robustness to market noise.
• Used a bottleneck layer in the autoencoder to enforce dimensionality reduction and feature learning.
• Transformed price prediction into a classification task by predicting price direction rather than exact values.
• Implemented triple barrier labeling to define exit points based on price targets (profit and stop-loss) rather than fixed horizons.
• Developed a new optimization metric specifically designed for triple barrier-labeled data to improve model training and evaluation.

Experimental results

Research questions

• RQ1 Does data augmentation and denoising via autoencoders improve the performance of a financial trading strategy?
• RQ2 Does triple barrier labeling improve classifier performance compared to simple direction classification?
• RQ3 Does hyperparameter tuning enhance the performance of the investment strategy?
• RQ4 Can a novel optimization metric be effectively derived and applied to triple barrier-labeled financial data?
• RQ5 How sensitive is the optimal strategy performance to variations in noise level and autoencoder bottleneck size?

Key findings

• Data augmentation with Gaussian noise and denoising via autoencoders significantly improved strategy performance, with the best results observed at balanced noise and bottleneck levels.
• Approach 3—using noise and autoencoder denoising—outperformed both baseline approaches across all bar lengths in terms of Information Ratio for an equally weighted portfolio.
• Triple barrier labeling outperformed simple direction classification for 15-minute and 30-minute bars, demonstrating better handling of market noise and improved optimization symmetry.
• The 5-minute bar performance with triple barrier labeling was inferior to the noise-augmented approach, indicating potential limitations in high-frequency settings.
• Hyperparameter tuning was essential, as optimal performance emerged only at a specific combination of noise level and bottleneck size, highlighting the non-linear relationship between these parameters and strategy returns.
• The study derived a novel optimization metric tailored for triple barrier labeling, enabling more effective training and evaluation of models under this labeling scheme.