Synthetic data in cryptocurrencies using generative models
| Authors | André Saimon S. Sousa et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2604.16182 |
| Download | |
| Categories | cs.LG, cs.AI |
Abstract
Data plays a fundamental role in consolidating markets, services, and products in the digital financial ecosystem. However, the use of real data, especially in the financial context, can lead to privacy risks and access restrictions, affecting institutions, research, and modeling processes. Although not all financial datasets present such limitations, this work proposes the use of deep learning techniques for generating synthetic data applied to cryptocurrency price time series. The approach is based on Conditional Generative Adversarial Networks (CGANs), combining an LSTM-type recurrent generator and an MLP discriminator to produce statistically consistent synthetic data. The experiments consider different crypto-assets and demonstrate that the model is capable of reproducing relevant temporal patterns, preserving market trends and dynamics. The generation of synthetic series through GANs is an efficient alternative for simulating financial data, showing potential for applications such as market behavior analysis and anomaly detection, with lower computational cost compared to more complex generative approaches.
Engineering Breakdown
Plain English
This paper addresses a critical problem in financial machine learning: real cryptocurrency data is often restricted due to privacy regulations and access limitations, yet researchers and institutions need large datasets to build accurate models. The authors propose using Conditional GANs with LSTM generators to synthesize realistic cryptocurrency price time series that maintain the statistical properties of real data without exposing sensitive information. The approach combines a recurrent neural network generator (LSTM-based) with a multilayer perceptron discriminator to adversarially train synthetic data generation. The method was validated across multiple crypto-assets and demonstrated the ability to reproduce relevant temporal and statistical patterns that would be useful for backtesting, research, and model development.
Core Technical Contribution
The core novelty is the application of conditional generative adversarial networks specifically architected for time series financial data, using an LSTM generator rather than standard convolutional or feedforward architectures. CGANs with conditional information allow the model to generate synthetic data conditioned on specific cryptocurrency features or market regimes, not just unconditional random samples. The combination of recurrent (LSTM) generation with adversarial training is designed to capture temporal dependencies and statistical consistency that standard GAN architectures miss in financial time series. This addresses a gap where most prior GAN work focused on image synthesis or non-temporal data, requiring domain-specific adaptation for the sequential, autocorrelated nature of price data.
How It Works
The system operates as a standard CGAN with two competing networks: (1) the generator takes random noise and conditioning information (such as asset type or market state) as input, then uses stacked LSTM layers to produce a synthetic time series of cryptocurrency prices with temporal coherence; (2) the discriminator is an MLP that receives generated or real sequences and learns to distinguish them while also respecting the conditioning information. During training, the generator learns to fool the discriminator by producing synthetic sequences that are statistically indistinguishable from real data while maintaining the same distributional properties and temporal patterns. The conditioning mechanism allows fine-grained control—you can generate synthetic data for specific cryptocurrencies or market conditions without retraining. The output is synthetic OHLCV (open-high-low-close-volume) time series or similar financial features that can be used directly in backtesting or model training pipelines.
Production Impact
For a financial institution or crypto trading firm, this solves the critical problem of obtaining large, unrestricted training datasets for algorithm development without exposing real trading data or customer information—a major regulatory and competitive advantage. Engineers would integrate this into data pipelines as a preprocessing/augmentation step: generate synthetic data from real historical data, then use the synthetic data for initial model validation, backtesting under edge cases, and sharing datasets across teams without privacy concerns. The practical benefit is faster iteration on strategy development and faster time-to-market for new trading signals, since you're not bottlenecked by data access agreements or compliance reviews. Trade-offs include: the synthetic data is only as good as the training data distribution (garbage in, garbage out), you need to validate that statistical moments and autocorrelation structure are truly preserved before relying on results, GPU compute for training the CGAN is non-trivial (likely hours on modern hardware), and mode collapse could cause the generator to produce limited diversity within each conditioning class.
Limitations and When Not to Use This
The paper does not address catastrophic failure modes like mode collapse (where the generator learns to produce only a narrow subset of realistic sequences), which is a known weakness of GANs—this could lead to synthetic data that underrepresents volatility spikes or rare market events. The approach assumes that the conditioning information is sufficient to capture market regimes; if truly novel market conditions emerge (black swan events, regulatory shocks), the generator will have no training signal for them and will likely produce invalid data. Validation against real data is mentioned but the abstract doesn't specify what statistical tests were used—in production, you'd need rigorous backtesting to confirm the synthetic data produces similar out-of-sample trading performance before you trust it for real decisions. The method requires paired real cryptocurrency data for each asset, so it doesn't generalize to newly-launched tokens or illiquid pairs with sparse history, limiting its applicability in rapidly-evolving crypto markets.
Research Context
This work builds on the established CGAN framework (Mirza & Osindski, 2014) and applies it to a domain (financial time series) where GANs have shown promise but face unique challenges around temporal coherence and statistical consistency. Prior work on synthetic financial data mostly used simpler methods (statistical resampling, parametric models) or lacked the temporal structure that LSTMs provide; this paper merges advances in sequence modeling (LSTMs) with adversarial learning for financial applications. The research connects to broader efforts in privacy-preserving machine learning and synthetic data generation, which is relevant for regulatory compliance (GDPR, data residency rules) and competitive advantage in finance. It opens directions for future work: using more sophisticated discriminators (e.g., Transformer-based or multi-scale), conditional mechanisms that respect market microstructure, and validation against downstream task performance (actual trading P&L) rather than just statistical metrics.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::