Data-driven methods for simulation and forecasting of financial time series

This thesis develops data-driven methods for the simulation and forecasting of financial time series. The contributions are structured into four main components. In the first part, we propose Tail-GAN, a novel nonparametric approach that combines a Generative Adversarial Network (GAN) with the joint elicitability property of Value-at-Risk (VaR) and Expected Shortfall (ES) for learning to simulate price scenarios that preserve tail risk features for a set of benchmark trading strategies. In the second part, we investigate the impact of order flow imbalance (OFI) on price movements in equity markets in a multi-asset setting. Our results show that, once the information from multiple levels is integrated into the OFI, multi-asset models with cross-impact do not provide additional explanatory power for contemporaneous impact compared to a sparse model without the cross-impact terms. We show however that cross-asset OFIs do improve the forecasting of future returns. In the third part, we apply machine learning models to forecast intraday realized volatility (RV), by exploiting commonality in intraday volatility by pooling stocks together, and by incorporating a proxy for market volatility. Neural networks dominate linear regression and tree-based models in terms of performance, and remain robust and competitive on unseen stocks not included in the training set, thus providing new empirical evidence for a universal volatility mechanism among stocks. We also propose a new approach to forecasting one-day-ahead RVs using past intraday RVs as predictors, and expose interesting time-of-day effects that aid the forecasting mechanism. In the last part, we develop a method for forecasting the realized covariance matrix of asset returns in the U.S. equity market by exploiting the predictive information of graphs in volatility and correlation. Specifically, we augment the Heterogeneous Autoregressive (HAR) model via neighborhood aggregation on these graphs. The results generally suggest that the augmented model incorporating graph information yields both statistically and economically significant improvements for out-of-sample performance over the traditional models.