Non-linear time-varying stochastic models for agroclimate risk assessment

Abstract

This work develops a model for minimum temperature in order to assess the weather related risk in agriculture industry. Non-linear autoregressive models with time-varying coefficients and volatility with various seasonal components and lags are compared in an appropriate model-selection algorithm using AIC. The optimal model is a time-varying autoregressive model which includes non-linear and seasonallyvarying autoregressive terms as well as time-varying volatility. These models are then used to simulate future weather from which the probabilities of appropriate complex hazard events are estimated.

A flexible forecasting model for production systems

“…The main reason is models which encapsulate the mean parameters and volatility parameters into one model often require slow Maximum Likelihood Estimation (e.g. in Hosseini et al (2015a)) or intensive Bayesian computation (e.g. West and Harrison (1997)) due to parametric form of the volatility.…”

“…Integrated models of course can be considered where the model includes all components (e.g. Hosseini et al (2015a) or West and Harrison (1997)).…”

“…However, we consider a two-component model for the following reason: (a) more flexibility in the mean components model in terms of features and algorithm; (b) more flexibility in the volatility model; (c) considerable speed gain by avoiding computationally heavy Maximum Likelihood Estimation (e.g. Hosseini et al (2015a)) or Monte Carlo Methods (e.g. West and Harrison (1997)).…”

“…This can indeed occur in the time series context for various reasons, including the divergence of the simulated values (see e.g. Hosseini et al (2015a)).…”

“…Some important examples include: Classical time series models such as ARIMA (e.g. Hyndman and Athanasopoulos (2014)) and GARCH (Tsay (2010)); Exponential Smoothing Based methods (see Winters (1960)); State-space models (see Kalman (1960), Durbin and Koopman (2012), West and Harrison (1997)); Generalized Linear Models extensions to time series (Kedem and Fokianos (2002), Hosseini et al (2015a)); Deep Learning based models such as LSTM (Hochreiter et al (1997)). Our framework utilizes the powerful aspects of these various models.…”