Interpretable GenAI: Synthetic Financial Time Series Generation with Probabilistic LSTM

This paper introduces a novel approach to generating synthetic financial market time series using a probabilistic Long Short-Term Memory (LSTM) neural network. We leverage generative AI to create data that replicates the probability distribution of real market data. This approach addresses the critical balance between model performance and interpretability, which is essential for practitioners in regulated industries. Our model enables the generation of realistic and diverse market scenarios that enhance risk management, scenario analysis, and trading strategy optimization. Key features include the ability to quantify and control the trade-off between interpretability and performance, and to reflect non-linear relationships, market regime changes and external variables. In a specific linear edge case, we show the capability to jointly train ARMA and GARCH components, each incorporating external variables, a feat that, to our knowledge, no existing Python library can achieve. For non-linear versions of the model, our experiments demonstrate superior performance compared to traditional ARMA-GARCH models, highlighting practical applications and benefits for financial institutions. This work advances effective use of generative AI in finance, offering a powerful tool for improving data privacy and expanding datasets, which can enhance model training.