Paper deep dive
Financial time series augmentation using transformer based GAN architecture
Andrzej Podobiński, Jarosław A. Chudziak
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/20/2026, 11:16:17 PM
Summary
This paper proposes a Transformer-based Generative Adversarial Network (TTS-GAN) to augment scarce financial time series data, specifically Bitcoin and S&P500 prices. The study demonstrates that training a Long Short-Term Memory (LSTM) forecasting model on data augmented by TTS-GAN significantly improves predictive accuracy compared to using real data alone. A novel quality metric, DTW DeD-iMs, combining Dynamic Time Warping and Deep Dataset Dissimilarity Measure, is introduced to monitor GAN training convergence and evaluate synthetic data quality.
Entities (9)
Relation Signals (6)
PyTorch → implements → TTS-GAN
confidence 95% · utilizing optimized PyTorch modules
LSTM → trainedonaugmenteddata → TTS-GAN
confidence 95% · 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
TTS-GAN → usedforaugmentation → Bitcoin
confidence 95% · We confirm these results across different financial time series (Bitcoin and S&P500 price data)... training a Long Short-Term Memory (LSTM) forecasting model on a dataset augmented with synthetic data generated by a transformer-based GAN (TTS-GAN)
TTS-GAN → usedforaugmentation → S&P500
confidence 95% · We confirm these results across different financial time series (Bitcoin and S&P500 price data)... training a Long Short-Term Memory (LSTM) forecasting model on a dataset augmented with synthetic data generated by a transformer-based GAN (TTS-GAN)
DTW DeD-iMs → evaluatesqualityof → TTS-GAN
confidence 90% · 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.
Simplified Gradient Penalty → enhances → TTS-GAN
confidence 88% · utilizing optimized PyTorch modules and a ”Simplified Gradient Penalty” to ensure stable convergence on volatile data.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2602.17865v1
- Canonical: https://arxiv.org/abs/2602.17865v1
Trouble viewing inline? Open PDF directly →
Full Text
35,590 characters extracted from source content.
Expand or collapse full text
Financial time series augmentation using transformer based GAN architecture Andrzej Podobiński 1a , Jarosław A. Chudziak 1b 1 Faculty of Electronics and Information Technology, Warsaw University of Technology, Warsaw, Poland andrzej.podobinski.stud@pw.edu.pl, jaroslaw.chudziak@pw.edu.pl Keywords: deep learning, time series, augmentation, forecasting, machine learning, artificial intelligence, financial time-series Abstract: 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. 1 INTRODUCTION Time-series forecasting is a foundational and crit- ical task that helps in high-stakes operational decisions across diverse domains, from optimiz- ing energy grids to predicting financial mar- ket volatility (Hyndman and Athanasopoulos, 2018). In recent years, deep learning mod- els have emerged as the state-of-the-art method for these complex challenges, often outperform- ing traditional statistical approaches and setting new benchmarks for the forecasting task (Lu et al., 2025). This progress is especially impor- tant within financial time-series analysis, where even marginal improvements in forecasting accu- racy translate directly into significant profitabil- a https://orcid.org/0009-0003-7574-2922 b https://orcid.org/0000-0003-4534-8652 ity gains. Despite their performance, deep learning mod- els are inherently data-hungry. Their perfor- mance is highly dependent on dataset size, and training them with insufficient data severely lim- its their utility (Bhatt et al., 2024). This is- sue is acutely pronounced in domains character- ized by scarcity and high dynamism, such as fi- nancial markets. Financial time series are in- herently volatile and characterized by evolving, non-stationary market regimes (Mills and Markel- los, 2008), as demonstrated in Fig. 1. This re- sults in a short time horizons of data relevant for model training. When faced with such limited data, complex deep learning architectures become prone to overfitting (Kaplan et al., 2020). The key question this research aims to ad- dress is: Can a transformer based GAN be suc- arXiv:2602.17865v1 [cs.LG] 19 Feb 2026 Figure 1: Bitcoin price and volatility before and after US president election 2024 cessfully used for data augmentation in order to improve the generalization and predictive capa- bilities of deep learning models when applied to volatile financial time series? We hypothesize that a data augmentation using TTS-GAN archi- tecture, can effectively mitigate the limitations of data scarcity in financial time series by generat- ing synthetic samples that capture the underly- ing statistical and temporal properties of the real data. To verify this hypothesis, we present compre- hensive experiment that compares performance of a standard Long Short-Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997) forecasting model trained only on real data against the same model trained on an augmented dataset. We demonstrate that augmentation leads to a reduc- tion in the Mean Squared Error (MSE), validating the GAN approach for financial time series. Our contributions include the demonstration of a TTS-GAN as an effective data augmentation tool for overcoming scarcity in financial time se- ries. We introduce technical enhancements to the TTS-GAN architecture, specifically utilizing op- timized PyTorch modules and a ”Simplified Gra- dient Penalty”(Mescheder et al., 2018) to ensure stable convergence on volatile data. Addition- ally, we propose a quality metric, DTW DeD- iMs, which combines Dynamic Time Warping and Deep Dataset Dissimilarity Measure to reliably evaluate both the temporal fidelity and distribu- tional similarity of synthetic sequences. The remainder of this paper is organized as follows: Section 2 reviews related work regard- ing generative models for time series and data augmentation in the financial domain. Section 3 details our approach, including technical en- hancements to the TTS-GAN architecture and the introduction of the DTW DeD-iMs conver- gence metric. Section 4 describes the experimen- tal design and the datasets preparation. Section 5 presents and discusses the results of the fore- casting improvements , and Section 6 concludes the paper with a summary of findings and future research directions. 2 RELATED WORKS In the domain of deep learning for time series analysis, the scarcity of sufficiently large and representative datasets poses a significant chal- lenge, as highlighted by Wen et al. (Wen et al., 2020). Effective augmentation methods must not only generate diverse samples but also accurately mimic the underlying temporal dependencies and statistical characteristics of real-world data. In this section, we review different GAN architec- tures for time series found in the recent litera- ture and the state of augmentation in financial data. Additionally, we discuss challenges in GAN training and detail architectural and algorithmic enhancements for stable GAN convergence. 2.1 Generative Models for Time Series While GANs have garnered considerable atten- tion, the challenge of generating high-fidelity time series data is distinct from image or text gen- eration. Early works, such as Recurrent GAN (RCGAN) and TimeGAN (Yoon et al., 2019) , leveraged Recurrent Neural Networks (RNNs) to capture temporal dynamics. RNN-based ap- proaches, however, often struggle with long-range dependencies and maintaining coherence across extended time sequences due to the inherent lim- itations of standard RNNs (Esteban et al., 2017) (Pascanu et al., 2013). While TimeGAN (Yoon et al., 2019) partially mitigates this by incorpo- rating a supervised loss component, these archi- tectures are fundamentally constrained by the se- quential processing of RNNs, limiting their abil- ity to effectively model the complex, multi-scale dependencies characteristic of non-stationary fi- nancial data. The emergence of the Transformer architec- ture has provided an efficient mechanism for modeling these long-range dependencies through the self-attention mechanism. The TTS-CGAN (Transformer Time-Series Conditional GAN) (Li et al., 2022b), based on TTS-GAN (Li et al., 2022a) initially demonstrated on biosignal data, Figure 2: TTS-GAN architecture(Li et al., 2022a) adopts this strength by building its generator and discriminator components entirely on transform- ers. This architectural choice is a fundamental advantage over RNN-based GANs, as it allows the model to process the entire sequence in par- allel and more effectively capture the intricate, non-linear relationships and temporal dependen- cies in financial timeseries datasets. 2.2 Augmentation for financial datasets Research confirming the utility of data augmen- tation in the financial domain provides a strong foundation for our work. A key study by Fons et al. (Fons et al., 2020) alongside the com- prehensive survey by Iwana and Uchida (Iwana and Uchida, 2021) emphasizes that augmenta- tion is a powerful, model-independent strategy. The authors specifically address the issue of lim- ited training data, noting that augmentation is highly effective for reducing overfitting and en- hancing the generalization capabilities of neural networks. To establish their baseline, Fons et al. used an LSTM model and the S&P500 dataset, which makes their experimental context highly relevant to our research. Crucially, while these prior works validate the overall benefit of augmentation in time series and compares various methods, they do not include a key analysis relevant to our approach. Specif- ically, neither the research by Fons et al. (Fons et al., 2020) nor the research by Iwana and Uchida (Iwana and Uchida, 2021) provides results for augmentation performed using the GAN frame- work. Our study aims to address this gap by ex- ploring the impact of GAN-based augmentation on stock forecasting accuracy. 2.3 Model convergence and algorithmic enhancement One of the challenges pronounced in the litera- ture is the lack of consensus on metrics for eval- uating generated data. This issue is particu- larly visible in the time-series domain, which re- mains comparatively less explored (Brophy et al., 2023). Jeng et al. (Jeng, 2025) recently recog- nized similar challenges when applying GANs for the augmentation of time-series data in electric vehicle battery state-of-charge modeling. Tra- ditional measures, such as the Wasserstein dis- tance, assess the overall data distribution but fail to adequately penalize dissimilarities of tempo- ral ordering or phase alignment (Cohen et al., 2021). Our work directly addresses this gap by proposing a novel evaluation method that com- bines DTW (Berndt and Clifford, 1994) and DeD- iMs (Calderon-Ramirez et al., 2022). This com- bination creates a measurement method that can account for temporal similarity and overall data distribution. The choice of DTW for quantifying temporal similarity was based on multiple suc- cessful applications of this method over recent decades (Sakoe and Chiba, 2003)(Bartolini et al., 2005)(Kholmatov and Yanikoglu, 2005)(Prabhu et al., 2024)(Jeng, 2025). Furthermore, our contribution extends to re- solving an architectural limitation of the original TTS-GAN architecture (Li et al., 2022a). We im- prove the model by integrating a performance- optimized PyTorch-based Transformer block (Paszke et al., 2019). This change is coupled with the utilization of an enhanced training al- gorithm designed to achieve better convergence stability (Mescheder et al., 2018), thereby con- tributing to more reliable and high-quality syn- thetic time-series generation. 3 PROBLEM AND APPROACH Applying complex deep learning models to finan- cial time series is limited by data scarcity and non-stationarity of underlying data distributions (Mills and Markellos, 2008). Forecasting models, particularly ones based on artificial neural net- works like the LSTM used in this study, require large datasets to effectively learn robust tempo- ral dependencies and prevent overfitting (Kaplan et al., 2020). Since financial market properties change rapidly, data collected over long periods can become irrelevant. Our approach mitigates this by generating high-quality synthetic data, that faithfully reproduces the characteristics of the limited in size training sets. 3.1 The generative model We selected a TTS-GAN (Li et al., 2022a) as the generative model, leveraging the Transformer’s self-attention mechanism, which is better than RNNs for modeling the long-range temporal de- pendencies critical in financial data. We acknowledge that the original TTS-GAN architecture has been published elsewhere. How- ever, in order to adopt it for financial domain we contribute by introducing technical enhance- ments to the implementation and training algo- rithm necessary for achieving stable convergence: 1. Optimized Architecture Implementation: To achieve optimal computational efficiency and training stability, the core Transformer mod- ules (encoder and decoder blocks) were im- plemented using the highly-optimized native PyTorch (Paszke et al., 2019) implementation rather than a custom implementation from original paper(Li et al., 2022a). This substi- tution significantly improved convergence of the model. 2. Enhanced Training Algorithm: The adversar- ial training process was modified to ensure stable learning. Specifically, we improved the training process from the original TTS-GAN’s by adding a ”Simplified Gradient Penalty,” which is calculated only on real samples. This technique was introduced by Mescheder et al. (Mescheder et al., 2018). This turned out to be critical for stabilizing the adversarial bal- ance and accelerating the convergence rate of both the generator and discriminator. These targeted improvements address the practical challenge of successfully training a gen- erative architecture on limited, volatile data, re- sulting in the high-fidelity augmented dataset. 3.2 The forecasting model For the forecasting component, we intention- ally employed a standard LSTM network despite the availability of more advanced architectures (Kwiatkowski and Chudziak, 2025). This choice serves the primary research objective: isolating and quantifying the performance gain achieved solely through the GAN-based data augmenta- tion process. By focusing on the basic LSTM architecture, we provide a clear validation of the utility of the augmentation technique, independent of specific model-level architectural advantages. In addi- tion, the LSTM architecture is a popular choice for a reference forecasting model, e.g. applied in the research by Fons et al. (Fons et al., 2020), in the research on TimeGAN (Yoon et al., 2019), in the research by Bańka (Bańka and Chudziak, 2025) and many others. The choice of LSTM for our experiments makes the results of our work more relatable with the results of other research. 3.3 Hyper-parameter tuning In order to ensure high quality of outcomes, hyper-parameters of both GAN and LSTM mod- els were tuned. Hyper-parameter optimization for GAN model required tuning parameters of both generator and discriminator components. We optimized number of transformer encoder blocks (D g andD d ), embedding size (M g andM d ), num- ber of attention heads (H g andH d ) and patch size (P g andP d ). The optimization process yielded the following parameter values:D g =3,D d =3; H g =5,H d =30;M g =10,M d =90; andP g =15, P d =15. A key observation from these results is that the discriminator required significantly higher capacity (specifically in its embedding size,M d =90, and number of attention heads, H d =30) than the generator. For LSTM model we used PyTorch’s n.LSTM module (Paszke et al., 2019). The optimization process focused on three key hyper- parameters: hidden_size, num_layers, dropout. Since the hyper-parameter space was relatively small, we opted for a simple, exhaustive grid search to find the optimal configuration. This optimization process yielded the following values: a hidden size of 64, three recurrent layers, and a dropout rate of 0.2. 3.4 Monitoring the quality of generated data An important part of the augmentation process is ensuring that samples used to extend the size of the original dataset are realistic. The sam- ples should exhibit characteristics of the original data samples while not being identical. Achiev- ing this required proper monitoring of the model convergence. In this section, we describe the monitored metrics that allowed us to detect and overcome the inherent challenges associated with GAN training, such as mode collapse, vanishing gradients, and general instability (Megahed and Mohammed, 2024). Detecting these issues was re- quired to ensure the quality and diversity of the synthetic data necessary for effective augmenta- tion. External metrics are required to monitor training progress because the GAN loss functions are uninterpretable. The generator loss is cal- culated on the basis of the number of samples that the discriminator correctly classified. Dis- criminator loss is calculated on the basis of mis- classified samples. The loss of the generator is a function of the discriminator parameters and vice versa(Goodfellow et al., 2014) (equations 1 and 2). That property makes it impossible to qualify if the model is converging based on values of loss functions. d_loss=MSE(D(real),1)+MSE(D(G(z)),0)(1) g_loss=MSE(D(G(z)),1)(2) To measure the model convergence we used ex- ternal measures, well established Wasserstein dis- tance (Santambrogio, 2015) and proposed modi- fied version of DeD-iMs (Calderon-Ramirez et al., 2022) based on the DTW distance. Both of this methods can be used to measure the similarity between the generated synthetic dataset and the real dataset. 3.4.1 Proposed Dynamic Time Warping based Convergence Metric Measuring convergence with a metric based on Dynamic Time Warping (DTW) was necessitated by the inherent limitations of the Wasserstein dis- tance. While Wasserstein distance is applicable to time series, it primarily treats the data as prob- ability distributions, ignoring the crucial chrono- logical order and phase alignment of events that define sequential similarity (Cohen et al., 2021). DTW, an elastic dissimilarity measure, is there- fore better suited to account for the stretching, permutation, and shifting of important patterns between real and synthetic sequences. However, since we require a measure of similarity between entire datasets, we propose combining DTW with the Deep Dataset Dissimilarity Measure (DeD- iMs (Stolte et al., 2024)). We call this new mea- sure DTW DeD-iMs. We demonstrate this com- bined approach as a simple yet effective method for comparing time-series datasets, proving par- ticularly useful for monitoring the convergence of the GAN training process. The DTW (Sakoe and Chiba, 2003) (Berndt and Clifford, 1994) enables measuring distances between individual samples. The DTW is a well established method that was already tested in many practical applications, like similarities based querying of image databases (Bartolini et al., 2005), matching shapes for authentication purposes (Kholmatov and Yanikoglu, 2005) and many others. DTW is an elastic dissimilarity measure which works by minimizing the cost of alignment between two time series. The measure can be described with the following formulas: d DTW (x,y)=D M,N (x,y)(3) Wherexandyare time series ofMandN points accordingly andD M,N (x,y)is described as: D i,j (x,y)=f(x i ,y j )+min D i,j−1 (x,y) D i−1,j (x,y) D i−1,j−1 (x,y) (4) We usedf(x i ,y j )=(x i −y j ) 2 , which is the most common choice for a cost function. The Deep Dataset Dissimilarity Measure (Stolte et al., 2024)(Calderon-Ramirez et al., 2022) is the second component required to mea- sure similarity between datasets. By providing a framework for comparing entire sets, DeD-iMs complements the base distance metric—in this case, DTW—which measures similarity only be- tween individual samples. The process for mea- suring similarity between two sets,S a andS b , in- volves the following steps: 1. Draw a random sample of sizenfrom setsS a andS b :S k a andS k b . 2. Calculated i =min k ∥s i −s k ∥ d for each sample s i ∈S a , wheres k ∈S b is the closest element to s i ∈S a measured with measured. In our im- plementation the distance between each pair fromS a andS b was calculated using the DTW. This yields a list ofndistancesD d . Figure 3: Comparison between TTS-GAN conver- gence measured using Wasserstein distance and DTW DeDiMs metric 3. Calculate a reference list of distances for the same samples in the datasetS a from itself - D ′ d . 4. The value of the dissimilarity measurement is an average of absolute differences between ref- erence and inter-dataset distances. Figure 3 plots the values of the convergence metric, clearly showing that both metrics con- verge during training. However, when comparing this result to the convergence measured using the Wasserstein distance, a key difference emerges: the DTW DeD-iMs metric continues to decrease through the later epochs. This behavior suggests that relying on DTW enabled the detection of subtle improvements in the temporal fidelity of the generated sequences, even after the overall feature distribution had converged. Using Wasserstein distance and our proposed DTW DeD-iMs ensured high quality of data used for augmentation which was necessary to validate our core hypothesis. 4 EXPERIMENTAL DESIGN AND PROCEDURE The primary objective of this study is to quan- tify the impact of utilizing GAN-augmented data on the forecasting performance of a deep learn- ing model. We hypothesize that training on an augmented dataset will yield a lower generaliza- tion error compared to training exclusively on real data. To verify this, we propose a three-phase procedure: 1. Generative Model Training: A TTS-GAN is trained on real segments of the stock price time series to generate high-fidelity synthetic data. 2. Baseline Forecasting: The LSTM forecasting model is trained and evaluated using only real samples from the financial datasets. 3. Augmented Forecasting: The same LSTM model is trained and evaluated using an aug- mented dataset, which combines the original real samples with the synthetic samples gen- erated by the trained GAN. To account for the non-stationarity and regime shifts inherent in financial time series, the entire experiment was conducted on 40 distinct datasets, each comprising price data from differ- ent, non-overlapping time periods. This approach allows us to test the augmentation method’s gen- eralization ability across diverse market condi- tions. We used historical daily closing prices for Bitcoin (last 10 years) and the S&P500 In- dex (1990–2025), sourced from publicly available sources. 1. Preprocessing: Data were initially smoothed using a backward-looking exponential moving average to mitigate the influence of transient outliers and high-frequency noise. 2. Sample Generation: The time series were divided into overlapping sequential samples, each consisting ofKtime points. We pre- pared datasets with sequence lengthsK=90 andK=120. The forecasting model predicts Sdata points based onTobservations, where K=T+S. ForK=90,Twas set to60and Sto30and forK=120,T=80andS=40. The resulting sample counts across the 40 time windows were: • 10x Bitcoin,K=90:≈370samples each • 10x Bitcoin,K=120:≈340samples each • 10x S&P500,K=90:≈890samples each • 10x S&P500,K=120:≈860samples each 3. Sample-wise Normalization: EachK-point sample was normalized independently using the MinMaxScaler transformation: X scaled = (X−X min ) (X max −X min ) (5) To prevent information leakage from the fu- ture, theX max andX min values were calculated solely based on the initial observation portion (Tdata points) of the sample, not the future forecast window (S). Each dataset was split into a training set (T), validation set (V), and test set (E) using the pro- portionsT:60%,V:20%, andE:20%. The GAN was trained onT∪Vand never shown any samples from the test setE. For the aug- mented experiments, we employed a1 : 1augmen- tation ratio, doubling the size of the training set (|T| augmented =2|T|). 5 RESULTS AND DISCUSSION To validate the hypothesis stated in this study, we measured the improvement in forecasting er- ror across 40 different datasets. We trained an LSTM on both original and augmented datasets, resulting in 40 pairs of forecasting models. For each pair, we calculated the MSE on the test set and computed the difference between the model trained on original data and the one trained on augmented data. Next, we grouped these results by dataset type (e.g., all Bitcoin datasets consist- ing of samples of length 90). The mean MSE im- provement was calculated, as well as the standard error (SE) and a paired t-test (p-value). Aggre- gated results are presented in Table 1. Across all four experimental conditions, aug- mentation led to a reduction in the average MSE compared to the baseline. Low Standard Error (SE) indicates that the sample mean is a highly reliable estimate of the true population mean. The consistently low SE values in our experiments suggest that measured results are reliable and re- producible. Specifically, a p-value below 0.05 in- dicates statistical significance of these results and confirms stated hypothesis. Larger improvements in MSE occurred for the Bitcoin dataset at both K=120 and K=90. Given that the Bitcoin dataset was smaller than the S&P500 data, this result highlights that aug- mentation provides more performance gains when original dataset is limited. While the MSE improvements may appear modest in absolute terms in some conditions, their consistency and reproducibility across 40 distinct time windows are highly significant. In the high-stakes domain of financial forecasting, even marginal, consistent improvements in pre- dictive accuracy can translate to substantial prac- tical advantages. Furthermore, the positive im- pact was more pronounced on the Bitcoin dataset, which is characterized by higher volatility and smaller effective dataset size compared to the S&P500. This outcome aligns with our theoret- ical premise: the more data-scarce and volatile the domain, the greater the benefit derived from high-fidelity data augmentation. The experimental design, which utilized 40 time windows for training GANs and LSTMs, was critical for implicitly testing the model’s resilience across various market regimes (e.g., bull vs. bear markets, periods of high vs. low volatility). While working on verifying our hypothesis we contribute with technical enhancements to the TTS-GAN implementation and training strategy. This leads to demonstrably faster and more sta- ble convergence compared to the original archi- tecture. On top of that, we address a critical challenge in GAN training, the lack of consensus on metrics for evaluating generated data specific to time series (Brophy et al., 2023) (Jeng, 2025). We introduce a novel dataset similarity mea- sure adapted for time series that combines Dy- namic Time Warping (DTW) (Berndt and Clif- ford, 1994), to assess temporal fidelity, and the Deep Dataset Dissimilarity Measure (DeD-iMs) (Calderon-Ramirez et al., 2022), to assess overall distribution similarity. The introduction of the combined DTW DeD- iMs metric provided a more signal for monitor- ing the GAN training progress. We observed that during training (referenced in Fig. 3), the Wasserstein distance converged and flattened pre- maturely, suggesting that the model had achieved distributional similarity but providing no further guidance on improving temporal quality. In con- trast, the DTW DeD-iMs metric continued to de- crease linearly over later epochs. The promising results from this empirical study warrant several directions for future explo- ration. While we focused on the LSTM to isolate the effect of augmentation, subsequent research should include more complex deep learning ar- chitectures, such as transformer-based architec- tures, which have recently received more atten- tion in financial forecasting field (Szydłowski and Chudziak, 2024). Given that these larger models are more data-demanding and prone to overfit- ting, the augmentation technique may potentially yield even greater performance gains. Finally, we believe that our research opens venues for more practical applications. One such application could be an adaptive system for mar- ket regime shift detection. A main focus would be to investigate the potential of the discrimina- tor component to detect shifts in the underlying domain, such as market regime changes. Success- ful application could enable the implementation of a fully automated, self-adapting machine learn- ing solution. In such a system, the discriminator would detect a regime change and trigger the re- training of the generator for data augmentation datasetsequence lengthMean MSE ImprovementSEp-value Bitcoin900.1020.0240.002 Bitcoin1200.1320.0160.003 SP500900.0370.010.006 SP5001200.0440.0180.007 Table 1: Forecasting MSE Improvements: Augmented vs. Real Dataset and, subsequently, the training of the forecasting model on augmented dataset. This application of the GAN leveraging unsupervised learning could lead to a system that operates with limited need for human intervention. 6 CONCLUSION The research successfully validated the use of GAN-based data augmentation for enhancing deep learning forecasting capabilities in finan- cial time series. We utilized an enhanced Transformer-based GAN (TTS-GAN), which in- cluded key implementation and algorithmic mod- ifications necessary for achieving stable conver- gence on volatile, non-stationary financial data. The core finding is that training a standard LSTM forecasting model on augmented data, generated by our enhanced TTS-GAN, reduces MSE across 40 distinct datasets and various fore- casting horizons. This validates the utility of our high-fidelity generative approach for mitigating the critical challenges of data scarcity and regime shifts in financial markets. Furthermore, we contributed by introducing a novel convergence monitoring metric that com- bines DTW and the DeD-iMs. This metric was shown to be highly effective at assessing both the distributional and temporal fidelity of synthetic sequences, providing crucial insight into the train- ing process that standard methods like Wasser- stein distance often fail to capture. REFERENCES Bańka, F. and Chudziak, J. A. (2025). Applying in- former for option pricing: A transformer-based approach. arXiv preprint arXiv:2506.05565. Bartolini, I., Ciaccia, P., and Patella, M. (2005). Warp: Accurate retrieval of shapes using phase of fourier descriptors and time warping distance. IEEE transactions on pattern analysis and ma- chine intelligence, 27(1):142–147. Berndt, D. J. and Clifford, J. (1994). Using dynamic time warping to find patterns in time series. In Proceedings of the 3rd international conference on knowledge discovery and data mining, pages 359–370. Bhatt, N., Bhatt, N., Prajapati, P., Sorathiya, V., Alshathri, S., and El-Shafai, W. (2024). A data- centric approach to improve performance of deep learning models. Scientific Reports, 14(1):22329. Brophy, E., Wang, Z., She, Q., and Ward, T. (2023). Generative adversarial networks in time series: A systematic literature review. ACM Computing Surveys, 55(10):1–31. Calderon-Ramirez, S. et al. (2022). Dataset similar- ity to assess semisupervised learning under dis- tribution mismatch. IEEE Trans. Artif. Intell., 4(2):282–291. Cohen, S., Luise, G., Terenin, A., Amos, B., and Deisenroth, M. (2021). Aligning time series on incomparable spaces. In International con- ference on artificial intelligence and statistics, pages 1036–1044. PMLR. Esteban, C., Hyland, S. L., and Rätsch, G. (2017). Real-valued (medical) time series generation with recurrent conditional gans. arXiv preprint arXiv:1706.02633. Fons, E., Dawson, P., Zeng, X.-j., Keane, J., and Iosifidis, A. (2020). Evaluating data augmenta- tion for financial time series classification. arXiv preprint arXiv:2010.15111. Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27. Hochreiter, S. and Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8):1735–1780. Hyndman, R. J. and Athanasopoulos, G. (2018). Forecasting: principles and practice. OTexts. Iwana, B. K. and Uchida, S. (2021). An empiri- cal survey of data augmentation for time series classification with neural networks. Plos one, 16(7):e0254841. Jeng, S.-L. (2025). Generative adversarial network for synthesizing multivariate time-series data in electric vehicle driving scenarios. Sensors, 25(3):749. Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361. Kholmatov, A. and Yanikoglu, B. (2005). Identity authentication using improved online signature verification method. Pattern recognition letters, 26(15):2400–2408. Kwiatkowski, J. and Chudziak, J. A. (2025). Com- paring transformer models for stock selection in quantitative trading. In International Confer- ence on Modeling Decisions for Artificial Intelli- gence, pages 239–250. Springer. Li, X., Metsis, V., Wang, H., and Ngu, A. H. H. (2022a). Tts-gan: A transformer-based time- series generative adversarial network. In Inter- national conference on artificial intelligence in medicine, pages 133–143. Springer. Li, X., Ngu, A. H. H., and Metsis, V. (2022b). Tts- cgan: A transformer time-series conditional gan for biosignal data augmentation. arXiv preprint arXiv:2206.13676. Lu, G., Ou, Y., Wang, Z., Qu, Y., Xia, Y., Tang, D., Kotenko, I., and Li, W. (2025). A sur- vey of deep learning for time series forecasting: Theories, datasets, and state-of-the-art tech- niques. Computers, Materials and Continua, 85(2):2403–2441. Megahed, M. and Mohammed, A. (2024). A com- prehensive review of generative adversarial net- works: Fundamentals, applications, and chal- lenges. Wiley Interdisciplinary Reviews: Com- putational Statistics, 16(1):e1629. Mescheder, L., Geiger, A., and Nowozin, S. (2018). Which training methods for gans do actually converge? In International conference on ma- chine learning, pages 3481–3490. PMLR. Mills, T. C. and Markellos, R. N. (2008). The econo- metric modelling of financial time series. Cam- bridge university press. Pascanu, R., Mikolov, T., and Bengio, Y. (2013). On the difficulty of training recurrent neural net- works. In International conference on machine learning, pages 1310–1318. Pmlr. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. (2019). Pytorch: An im- perative style, high-performance deep learning library. Advances in neural information process- ing systems, 32. Prabhu, H., Valadi, J., and Arjunan, P. (2024). Gen- erative adversarial network with soft-dynamic time warping and parallel reconstruction for energy time series anomaly detection. arXiv preprint arXiv:2402.14384. Sakoe, H. and Chiba, S. (2003). Dynamic program- ming algorithm optimization for spoken word recognition. IEEE transactions on acoustics, speech, and signal processing, 26(1):43–49. Santambrogio, F. (2015). Optimal transport for ap- plied mathematicians. Springer. Stolte, M., Kappenberg, F., Rahnenführer, J., and Bommert, A. (2024). Methods for quantifying dataset similarity: a review, taxonomy and com- parison. Statistic Surveys, 18:163–298. Szydłowski, K. and Chudziak, J. A. (2024). Hidformer: Transformer-style neural network in stock price forecasting. arXiv preprint arXiv:2412.19932. Wen, Q. et al. (2020). Time series data augmenta- tion for deep learning: A survey. arXiv preprint arXiv:2002.12478. Yoon, J., Jarrett, D., and Van der Schaar, M. (2019). Time-series generative adversarial networks. Ad- vances in neural information processing systems, 32.