A0707
Title: Robust hidden semi-Markov models for meteorological residuals in the Venice lagoon
Authors: Lorena Ricciotti - University of Bari (Italy) [presenting]
Alfonso Russo - University of Rome Tor Vergata (Italy)
Sondre Hoelleland - Norwegian School of Economics (Norway)
Antonello Maruotti - Libera Università Maria Ss Assunta (Italy)
Abstract: The aim is to present a flexible statistical framework to analyze the tides' meteorological component in the Venice lagoon. A robust class of hidden semi-Markov regression models (HSMRMs) is proposed, capturing key marine data features such as regime switching, time-varying heteroskedasticity, heavy tails, skewness, and outliers. The framework extends hidden Markov regression models, relaxing the geometric sojourn time assumption and incorporating variable selection via elastic-net regularization. To improve robustness to outliers and skewed data, Gaussian, Student-t, and Johnson's SU distributions are used, enabling more accurate modeling of sea level behavior. Empirical analysis is conducted using hourly data from the Lido di Porto inlet tide gauge, covering meteorological variables as wind speed and direction, air and water temperature, pressure, and humidity. The proposed model identifies four environmental regimes influencing meteorological residuals associated with specific weather conditions and temporal dynamics. The model distinguishes between true and apparent contagion by incorporating autoregressive components, thus addressing latent regime shifts and explicit temporal dependencies. Simulation studies confirm the ability of the Johnsons SU-based HSMRM in parameter estimation accuracy and classification performance under skewed data generation processes. The regularization approach effectively selects relevant covariates and lags, enhancing model interpretability.
