Financial time series augmentation using transformer based GAN architecture

Time-series forecasting is a critical task across many domains, from engineering to economics, where accurate predictions drive strategic decisions. However, applying advanced deep learning models in challenging, volatile domains like finance is difficult due to the inherent limitation and dynamic nature of financial time series data. This scarcity often results in sub-optimal model training and poor generalization. The fundamental challenge lies in determining how to reliably augment scarce financial time series data to enhance the predictive accuracy of deep learning forecasting models. Our main contribution is a demonstration of how Generative Adversarial Networks (GANs) can effectively serve as a data augmentation tool to overcome data scarcity in the financial domain. Specifically, we show that training a Long Short-Term Memory (LSTM) forecasting model on a dataset augmented with synthetic data generated by a transformer-based GAN (TTS-GAN) significantly improves the forecasting accuracy compared to using real data alone. We confirm these results across different financial time series (Bitcoin and S&P500 price data) and various forecasting horizons. Furthermore, we propose a novel, time series specific quality metric that combines Dynamic Time Warping (DTW) and a modified Deep Dataset Dissimilarity Measure (DeD-iMs) to reliably monitor the training progress and evaluate the quality of the generated data. These findings provide compelling evidence for the benefits of GAN-based data augmentation in enhancing financial predictive capabilities.

💡 Research Summary

This paper tackles the chronic problem of data scarcity in volatile financial time‑series forecasting, where deep learning models such as LSTMs often overfit due to limited, non‑stationary training samples. The authors propose to use a transformer‑based Generative Adversarial Network (TTS‑GAN) as a data‑augmentation engine, generating synthetic price sequences for Bitcoin and the S&P 500 that preserve the statistical and temporal characteristics of the original series.

Key technical contributions include: (1) a re‑implementation of the original TTS‑GAN architecture using native PyTorch transformer encoder‑decoder blocks, which improves computational efficiency and training stability; (2) the integration of the “Simplified Gradient Penalty” (Mescheder et al., 2018) applied only to real samples, thereby balancing the generator‑discriminator dynamics and accelerating convergence; (3) an extensive hyper‑parameter search that reveals a markedly larger discriminator (embedding size = 90, attention heads = 30) than generator, a design choice that helps the discriminator capture the complex distribution of financial returns.

To monitor the quality of generated series, the authors introduce a novel metric, DTW‑DeD‑iMs, which fuses Dynamic Time Warping (DTW) – an elastic distance that respects temporal alignment – with the Deep Dataset Dissimilarity Measure (DeD‑iMs), which quantifies distributional divergence at the dataset level. This combined metric overcomes the limitations of Wasserstein distance, which ignores ordering, and provides a practical early‑stopping signal during GAN training.

Experimental methodology: the authors train a baseline LSTM (hidden size = 64, 3 layers, dropout = 0.2) on raw data only, and a second LSTM on an augmented dataset that mixes real and synthetic sequences. Forecast horizons of 1, 5, and 10 days are evaluated for both Bitcoin and S&P 500 price series. Performance is measured by Mean Squared Error (MSE) and by the proposed DTW‑DeD‑iMs score.

Results: Across all horizons, the augmented LSTM achieves a reduction in MSE ranging from roughly 10 % to 15 % relative to the baseline, with the most pronounced gains on the highly volatile Bitcoin series. Moreover, lower DTW‑DeD‑iMs scores (indicating higher similarity between synthetic and real data) correlate strongly with larger MSE improvements, validating the metric’s usefulness as a quality indicator.

The paper also discusses limitations: the impact of varying the proportion of synthetic data is not explored, and comparisons with other forecasting architectures (e.g., Transformer‑Encoder, N‑BEATS) are absent. Future work is suggested in the directions of multi‑asset portfolio augmentation, reinforcement‑learning‑driven time‑series generation, and real‑time trading strategy integration.

In summary, the study demonstrates that transformer‑based GANs can effectively augment scarce financial time‑series, leading to measurable gains in downstream forecasting accuracy. The DTW‑DeD‑iMs metric provides a robust, interpretable tool for assessing synthetic data quality, addressing a notable gap in the evaluation of time‑series GANs. This work therefore offers a practical pathway for deploying deep learning models in data‑limited, high‑volatility financial environments.