Synthetic data generation with deep generative models to enhance predictive tasks in trading strategies

Abstract

This work develops machine learning (ML) predictive models on price signals for financial instruments and their integration into trading strategies. In general, ML models have been shown powerful when trained with large amounts of data. In practice, the time-series nature of financial datasets limits the effective amount of data available to train, validate and retrain models since special care must be taken not to include future data in any way. In this setting, we develop deep generative models to produce synthetic time-series data, enhancing the amount of data available for training predictive models. Synthetic data obtained this way replicates the distribution properties of real historical data, leads to better performance, and enables thorough validation of predictive models for price signals. We leverage machine-generated predictive signals on synthetic data to build trading strategies. We show consistent improvement leading up to profits in our simulations for commodities and forex exchange markets.