In food retail industry, there is a real need of daily sales forecast
to support the store managers in precise ordering, without compromising
food waste and stock-outs. However, the time series
which contain daily sales of perishable food in retail stores, are
usually characterized by high volatility and skewness, which are
also time varying. These are important constraints, but often
ignored in forecasting. In order to overcome these issues, there is a
need to develop a time series forecasting model which incorporates
uncertainty in forecasts and influence of external variables
such day-of-the-week seasonality, month-of-the year seasonality,
holidays, festivals, price reduction and weather on the sales. In this
study, SARIMA-MLR and SARIMA-QR models are developed and
applied to forecast the daily sales of banana in a German retail
store. Both of these models yield better predictions for out-sample
data, compared to seasonal naïve forecasting, SARIMA, and MLPNN
models. Other than this, the derivation of inventory policy from
the estimated sales forecast is one of the important problems in
stock management in food retail industry. The SARIMA-MLR model
produces only the point forecast, i.e., the mean forecast. As the
true distribution of demand or sales is not normal, the estimation
of prediction intervals from the point forecast is not going to
reflect the reality. However, the SARIMA-QR model has additional
benefits over the SARIMA-MLR model as follows: (i) it helps to
forecast the higher service levels directly and accurately without
extrapolation, (ii) the QR results offer detailed and focused insights
into the effects of the covariates, and (iii) when the focus of
interest is on the higher (for promotional activities) or lower sales
(due to extreme weather conditions), the model can aid the
management to make accurate and proper decisions.