Data Augmentation for Financial Machine Learning using Artificial Market Simulation: Application to Volatility Forecasting

Abstract

We show that artificial data from artificial market simulations are effective for training financial machine learning models. Financial time-series data is not suitable for training machine learning models due to limited amount of data available and the demand to use historical data for performance evaluation. On the other hand, the artificial market is the simulation model to understand macro-level phenomena through the designing of individual investors. By using the artificial market appropriately, any amount of various data can be generated. Our study show that machine learning models trained on artificial data outperform models trained only on real data. Specifically, we first propose aFCNAgent that reproduces the stylized facts of any stock markets, such as asymmetry of volatility and correlation between trading volume and volatility. Next, we show that the artificial data generated by the agent improves the performance of volatility prediction.