Spatiotemporal ARCH models capture temporal volatility persistence and cross-sectional dependence but typically require a predefined spatial weight matrix. This is restrictive in financial markets, where the dependence network is rarely known. We develop a LASSO-penalised quasi-maximum likelihood estimator that jointly learns a sparse weight matrix and estimates temporal dependence and covariate effects. Monte Carlo experiments show that the method recovers the model parameters and underlying network, with accuracy improving as the temporal sample size increases. We apply the method to daily returns from twenty UK-listed firms and compare the learned network with Euclidean-distance, correlation, autoregressive-similarity and sector-based structures. The learned network improves out-of-sample volatility prediction and reveals directional firm-level and cross-sector dependence not captured by the predefined alternatives. The method provides a data-driven framework for learning interpretable conditional-volatility networks.
🌊
