A1091
Title: Sparse causal dynamic linear regression
Authors: Rui Huang - Nanjing University (China) [presenting]
Abstract: Longitudinal datasets with multiple time series are common in various fields, such as environmental studies, economics, and finance, motivating the need for effective analytical methods. The dynamic regression model is a powerful tool for exploring linear temporal relationships between variables but may face numerical instability and intractability when handling long time series with multiple leads and lags using the time-domain approach. While the frequency domain approach provides an elegant and efficient alternative solution, the resulting model may suffer from non-causality and non-sparsity, hindering practical utility and interpretability. A novel sparse causal modification is proposed to the general dynamic linear regression model, formulated as a frequency domain functional optimization problem to address these challenges. An accelerated functional proximal gradient descent algorithm is then derived to numerically compute the solution. To further alleviate the computational burden, a computationally efficient one-pass estimation procedure that produces sufficiently good approximate solutions is also proposed when the true underlying model is approximately causal. The theoretical properties of the proposed algorithms are established. The efficacy of the proposed method with both simulation studies and a set of stock index return time series data is showcased.
