Supervised Autoencoder MLP for Financial Time Series Forecasting

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.

• The paper introduces a Supervised Autoencoder MLP with triple barrier labeling to refine price prediction in algorithmic trading.
• It employs noise-based data augmentation and hyperparameter optimization to improve risk-adjusted returns on S&P 500, EUR/USD, and BTC/USD.
• Results indicate the SAE-TBL approach outperforms traditional strategies, emphasizing critical tuning of noise levels and bottleneck dimensions.

Enhancing Algorithmic Trading Strategies with Supervised Autoencoder MLPs and Generative AI

Introduction

Algorithmic trading strategies often sift through high-frequency financial data seeking patterns that predict future price movements. Traditional methods rely heavily on daily closing prices, potentially overlooking intraday fluctuations that could inform more nuanced strategies. This study, embarked on by Bartosz Bieganowski and Robert Ślepaczuk from the University of Warsaw, aims to close this gap through the application of Supervised Autoencoder Multi-Layer Perceptrons (SAE-MLPs). By incorporating data augmentation via noise addition and employing triple barrier labeling, the study seeks to refine financial time series forecasting for the S&P 500, EUR/USD, and BTC/USD markets.

Methodology

The methodology centers around two principal components: the SAE-MLP model and triple barrier labeling. The SAE-MLP combines supervised learning with an autoencoding structure to predict future price movements. In contrast to simple price direction classification, it transforms return estimation issues into more granular, actionable output. The triple barrier labeling, meanwhile, estimates price actions not just based on direction but also on specified profit-taking and loss-stopping levels, creating a more nuanced prediction method.

The dataset spans from January 1, 2010, to April 30, 2022, segmented into training and testing phases. Models were tuned through hyperparameter optimization and evaluated using a walk-forward validation approach to mirror a real-world, evolvable strategy. The study's effectiveness was gauged on risk-adjusted return metrics, namely the Sharpe and Information Ratios.

Results

The study comprehensively compared four distinct neural network modeling approaches across three assets and an equally weighted portfolio. Approach 4, which combines SAE with Triple Barrier Labeling (SAE-TBL), emerged as notably effective under certain configurations, particularly for EUR/USD and SPX trading. This configuration outperformed the traditional buy-and-hold strategy when considering both the Sharpe Ratio and specific adaptations to combat overfitting and market noise interference.

Sensitivity Analysis

Sensitivity checks revealed a pronounced dependency of model performance on the data augmentation parameters and the size of the SAE's bottleneck. A sweet spot was identified for noise levels at approximately 5\% of annualized volatility, and a bottleneck size constituting 40\% of the original features. Deviations from these levels generally led to diminished returns, underscoring the importance of careful parameter calibration.

Conclusions and Implications

Findings suggest that the integration of SAE and TBL significantly elevates the potential of financial time series prediction models, enhancing risk-adjusted returns beyond traditional techniques. However, the study underscores the necessity for fine-tuned hyperparameter optimization to avert overfitting and ensure the strategy remains responsive to market dynamics.

The utilization of such advanced machine learning strategies in financial markets implies a shift toward better informed, more strategic trading approaches, potentially contributing to greater market efficiency. For regulators and financial institutions, these findings herald the need to adapt to an increasingly algorithm-driven trading landscape, ensuring that regulations foster innovation while safeguarding market integrity.

Future Research Directions

The study calls for further exploration into other forms of data augmentation, the impact of including transaction cost models, and various autoencoder architectures to potentially uncover even more effective strategies. Additionally, examining these methodologies across alternative financial instruments could validate their generalizability or indicate necessary adjustments.

In sum, the research presented by Bieganowski and Ślepaczuk offers a compelling case for the continued exploration and application of SAE-MLPs and generative AI in algorithmic trading, potentially setting the stage for the next leap in predictive financial modeling.