Comparison of Linear and Nonlinear Models for Forecasting of Food Commodity Prices in Latvia

22nd International Scientific Conference "Engineering for Rural Development" : Proceedings, Vol.22 2023
Oksana Pavļenko, Andrejs Matvejevs

Analysing real-life data of commodity price dynamics is challenging, there can be non-stationary, nonlinear, contain structural breaks. In this paper, we explore whether threshold models are preferable to linear autoregressive models (ARIMA) and whether the logistic smooth transition (LSTAR) model is preferable to the self-exciting threshold autoregressive (SETAR) model for important Latvian food commodity prices. Using historical prices of 16 most popular food products in Latvia over the last 18 years, we assess the goodness of fit of ARIMA (SARIMA), SETAR and LSTAR for each of the most popular commodity prices in Latvia and then compare the out-of-sample forecasts using measures RMSE and MAPE. Although different types of models appear to be most suitable for different commodities, even despite their similarity like fresh pork, chicken and beef, the overarching conclusion is that regime-switching models fit the prices of the majority of products better. ARIMA is the preferred model for some goods for construction of out-of-sample forecasts marginally more often than for the goodness of fit. Nevertheless, threshold models still appear superior in most cases. Additionally, we obtain rather large smoothness coefficients for most LSTAR models, which means that there are no significant reasons to prefer LSTAR to SETAR.

Keywords
time series, goodness of fit, out-of-sample forecast, structural break, ARIMA, SARIMA, LSTAR, SETAR
DOI
10.22616/ERDev.2023.22.TF145
Hyperlink
https://www.tf.lbtu.lv/conference/proceedings2023/

Pavļenko, O., Matvejevs, A. Comparison of Linear and Nonlinear Models for Forecasting of Food Commodity Prices in Latvia. In: 22nd International Scientific Conference "Engineering for Rural Development" : Proceedings, Vol.22, Latvia, Jelgava, 24-26 May, 2023. Jelgava: Latvia University of Life Sciences and Technologies, 2023, pp.735-744. ISSN 1691-5976. Available from: doi:10.22616/ERDev.2023.22.TF145

Publication language
English (en)