a. Problem Description
Our client is a German drug manufacturer who owns and operates six stores in different parts of
Germany. Each store has a unique layout and some stores are open for fewer days than others. Moreover, customer footfall varies by location. Our client contracts staffing personnel on a daily basis from a staffing agency for all the stores, and each contracted staff is paid on per hour basis every day. The
management in our client’s organization has been tasked with an objective to bring down the staffing
costs by optimizing the number of store staff that is contracted every day. The number of staff required at
each store depends on the number of customers who visit the store. Therefore, as a first step towards
optimizing the number of staff required at each of the six stores, our client wanted to estimate customer
footfall at each store.
Our task was to forecast customer footfall on a daily basis at each store location over a forecast period of
six weeks starting August 1, 2015. To accomplish this task, our client provided daily sales and customer
footfall information from January 1, 2013 to July 31, 2015.
b. Brief Description of Data
The dataset was obtained from Kaggle.com. For each store, the dataset contained daily sales and
customer footfall. In addition to these two fields, there was further information provided on whether there
was any sales promotion on any given day and whether a given day was a state holiday or a school
For the purpose of this analysis, only customer footfall was considered as it would directly impact the
staffing requirement at the store. The customer data contained level, noise and seasonality while flat trend
was observed for all the series. Seasonality was observed to be 7.
c. High Level Description of Forecasting Methods
The following methods were used for forecasting-
- Naïve (Benchmark Prediciton)
- Holt-Winters Smoothing – Due to the presence of seasonality
- Multi Linear Regression (MLR) – Due to the presence of seasonality
- Ensemble – To evaluate if the combination of model is better than individual models
The methods and results were evaluated based on Root Mean Square Error (RMSE) and based on the
plots of the actual and forecasted values from various methods.
d. Conclusions and Recommendations
- Based on the plots and errors above we conclude that Holt Winter’s method should be used for
forecasting for all stores.
- We believe that hourly data may help the store better predict the number of customer during peak
hours and this could enable the store managers to plan staffing appropriately while reducing the
total operational cost of the store.
- We also recommend that the store managers consider the 95% confidence band for staffing as
that account for majority of the cases rather than the single value of the forecast. This would
ensure that customer service levels are maintained across the stores.