Problem Description - Representing the hypermarket, the objective of our forecasting is to reduce the spoilage of vegetables in the hypermarket by accurately forecasting sales on a daily basis. By using historical data of the last year we plan to forecast daily demand for two SKU's 'Exotic Vegetables' and 'Beans' under the vegetable sub-class.
Since the two selected vegetable SKU's had a profit margin of close to 25% we tried to model the costs of under-prediction or over-prediction as an important metric to assess the performance of our model
Model Description - Final model used for our analysis is comprised of two steps and is a combination of two forms of forecasting method. We ran the Holt-Winter's Model on the data following it by a Multiple Linear Regression using the forecasts and the relevant holidays which do not fall on a weekend. We have used a Holt Winter's with no trend in the initial step as the visualizations of the raw data do not highlight any particular trend. However we have captured weekly seasonality where we observed that sales were highest on a Sunday every week followed by Saturday and Wednesday.
Model Performance - Since we had considered the Naive forecasts as a benchmark for our prediction we have taken and compared the RMSE and MAE parameters for our model obtained by conditioning the cost factor into our model with the Naive forecast models.
Forecasts and their assumptions - We generated one month forecast in the future to generate a daily forecast of vegetable supplies for the next 30 days. We planned to ensure that this model was updated every month by factoring in the errors that affected the model last month
Conclusions and Recommendations - We can fairly conclude that the model developed by us is effective to forecast daily vegetable sales for the next 30 days for the two vegetable SKU's. We observe that certain forecasts in July and Aug are not captured well. We believe this is the effect of a special event such as 'Rains' in the city disrupting normal events. We recommend that the predictive ability of the model can be improved by modeling in daily rainfall data.