The percentage of the
observed values that lie below the lower prediction interval is specified in the lower left
corner (1.6%). This same procedure is used for the out- sample data to compare the prediction
results SARIMA-QR and SARIMA-MLR models. The prediction intervals (5% and 95%) are estimated from
the point forecasts of SARIMA-MLR model using Eq. (3) suggested by Chatfield (2000) and Hyndman
and Athana- sopoulos (2013). The prediction intervals estimated from SARIMA-MLR model are plotted against the observed out-sample values of
sales of banana in Fig. 8. As shown in this figure, "' 5.5% of the observed values lie
outside the prediction intervals. Likewise, Fig. 9 shows the prediction intervals from
SARIMA-QR model that are plotted against the observed values of sales of banana for the
out-sample data. In this figure, the percentage of observed values that lie outside the prediction
intervals is around 3%. From this result, it is evident that the prediction intervals of
SARIMA-QR model can capture more than the theoretically defined uncertainty (95%). On the other
hand, the prediction intervals of SARIMA-MLR model covers nearly 94.5% of the out-sample
observations. Unlike the prediction intervals from SARIMA-QR model, it overestimates and
underestimates the upper and lower prediction intervals as shown in Fig. 8. In food retail
industry, the overestimation and underestimation of prediction intervals will definitely lead
to stock-outs and food waste.
The estimators of different quantiles can also be combined to form
a robust point estimator. In literature, it was described how to combine the estimators of
different quantiles using Tukey's trimean, Gastwirth estimator and five-quantile estimator (Koenker
and Bassett,
1978; Taylor, 2007). The robust point estimators are calculated using
the above mentioned methods and the results are shown in Table 6.