Multi-step forecasting of the Philippine storm frequencies using Poisson neural network

The paper aims to forecast the Philippine storm frequencies using nonlinear Poisson autoregressive model with exogenous variables. nonlinearity is defined by its kernel which an artificial neural network (ANN) one hidden layer and two output neurons, trained simultaneously semesters ahead for a given input. Furthermore, covariates studied were Average Sea Surface Temperatures in NINO3.4 region (5∘− 5∘S,170∘− 120∘W) eastern pole (0∘− 10∘S,90∘− 110∘E) of Dipole Mode Index. data, taken from Japan Meteorological Agency’s Regional Specialized Center time points running 1950 October 2021, modeled at semester-level granularity. estimation done maximum likelihood minimizing negative log-likelihood function. Bayesian hyper-parameter optimization was used tune across different activation functions, number training optimizers, train validation splits. Lastly, best then compared univariate (and counterpart) Negative Binomial Autoregressive model. proposed captures well characteristics data both terms point associated uncertainties.