Abstract:
Coconut is a perennial crop which significantly contributes to the Sri Lankan economy. It is the
third largest foreign exchange earner for the country. Recently Sri Lanka faced high reduction
in coconut yield. Coconut prices have doubled due to the shortage. A crop shortfall and a
drought have forced the country to import coconuts. The lack of the advanced models to forecast
the coconut prices also may be a reason for this crisis because the industry may have not had a
real idea of the pattern of the price changes. In this study two types of nonlinear time series
models were fitted to retail and producer price series, separately. Finally, out of the fitted
models, the best one was selected to forecast monthly producer and retail price of coconut.
Monthly producer and retail price series from 1994 to 2010, showed a non-constant variance
with the time. Since the Box Peirce LM test statistically confirmed that the volatility clusters are
present in each return series, Generalized Auto Regressive Conditional Heteroskedasticity
(GARCH) models were fitted to both return series. Monthly coconut production was taken as an
exogenous variable since cross correlograms indicate a significant correlation between each
price and production. As an alternative nonlinear approach Nonlinear Autoregressive neural
network with exogenous inputs (NARX) models were fitted for each price series considering
production as an exogenous variable. Forecasts of producer price and retail price of coconut
were obtained from each model. Since AIC (Akaike Information Criterion) values and MSE
(Mean Square Error) values were in different scales among the models, Mean Absolute
Percentage Error (MAPE) of prediction was used as the measure to compare the models. The
results showed that the NARX model was the most appropriate model to forecast monthly
producer price of coconut as well as retail price of coconut during the study period, because the
smallest MAPE of prediction was given by the NARX model.
