INPUT DATA RANGE OPTIMIZATION FOR FREIGHT RATE FORECASTING USING THE ROLLING WINDOW TESTING PROCEDURE

Abstract

This paper investigates the impact of sample size (input range) in predictive accuracy for fuzzy time series and autoregressive integrated moving average methodologies. The argument of this paper is the existence of an optimum sample size subject to out of sample forecasting accuracy. This phenomenon opposes to the common belief that larger sample size would result in more accurate predictions. A series of simulations are conducted to demonstrate the phenomenon explicitly to prove its impact. Empirical results clearly indicate the oscillations and possible existence of an optimum sample size for given algorithms. Although these two approaches are tested in the empirical study, results significantly emphasize possible existence of sample size asymmetries in other kinds of algorithms. For illustration of the phenomenon, Baltic Dry Index (BDI) is utilized in empirical simulations.