We investigate the predictive abilities of the heterogeneous autoregressivemodel compared to machine learning techniques across an unprecedented dataset of 1,455 stocks. Our analysis focuses on the role of fitting schemes, particularly the training window and re-estimation frequency, in determining the HAR model’s performance. Despite extensive hyperparameter tuning, ML models fail to surpass the linear benchmark set by HAR when utilizing a refined fitting approach for the latter. This study suggest that properly fitted HAR models provide superior forecasting accuracy, establishing robust guidelines for their practical application and use as a benchmark.

We propose a method using temporal convolutional networks for model parameter estimation by learning the mapping from sample data to its generating parameters. This map is then used for defining moment conditions in simulation-based inference. Our approach outperforms maximum likelihood estimators for small and moderate samples and effectively estimates a jump-diffusion financial model.

We conducted a Monte Carlo simulation study on the application of deep reinforcement learning methods for portfolio management with predictable returns and costly transactions. Our findings reveal varying algorithm performance, emphasizing the need to restrict asset samples for sufficient DRL performance. We also contributed to the Julia programming language ecosystem by providing codes for work replication and extension.

Using causal machine learning, we examine the influence of sentiment from earnings-related news articles on firms’ return, volatility, and trade volume. Our analysis considers price and volume reactions to different sentiments within varying economic, financial, and aggregated investor mood conditions. The findings indicate significant differences in the effects of sentiment types, larger reactions to negative sentiment, and investors’ general underreaction to news, especially in adverse macroeconomic conditions or high stocks liquidity.