Optimizing Operational Spend by Predicting Product Sales

Project Details




Aditya Verma, Amit Kumar Gupta, Neeraj Nathany, Prateek Singhvi, Varsha Shridhar, Vishal Abraham





In the retail business, it is crucial for the firms to accurately forecast sales in the future to
prepare themselves and optimize costs. Over-estimating sales can lead to a significant cost of inventory holding and even losses due to expiry in case of perishable items. There could also be a scenario of underutilized resources in manufacturing be it machinery or labour.
Underestimation of sales in forecasts can lead to loss of business opportunity.

Problem description
There are three main problems of the retailers that we are looking to solve

  • Allocating budget towards increasing manufacturing or production capacity
  • Creating scheduling plans for inventory holding of 2 particular stores to account for peak periods and lean periods
  • To calculate the required number of counters that need to be operational on a given weekday and the manpower required to monitor them

Data description

  • We have 5 years (2013 onwards) of time series data at daily level for sales of 50 products at 10 store locations
  • The dataset has 9,13,000 records and consists of categorical variables such as store id and product id and continuous variables such as sales
  • The data has been obtained from Kaggle.com where it was used as a part of a competition
  • There is no indication of the store names, locations, product name or product category. They have been hidden by assigning numbers to the names.
  • Hyperlink of the dataset: https://www.kaggle.com/c/demand-forecasting-kernels-only/data

Forecasting methods and performance evaluation
Following forecasting methods were used:

  • Linear regression
  • Holt-winters Smoothing
  • Naïve forecast

We looked at RMSE and MAPE values to evaluate the performance of forecasting models. The benchmarking was done against Naïve model which was assumed to be the status quo.

Conclusions and recommendations

  • Daily forecast for the 2 stores will be helpful in (a) staffing number of employees on feet, and (b) planning number of counters to be open on which day of the week

As can be seen from the data, there has been a consistent trend in sales during weekends and sales during weekdays. It is notable that the sales start to peak during weekends i.e., Saturday and achieve at its peak on Sunday followed by a sudden drop on the next day.​
Firm can hire contract employees for weekends – Saturday and Sunday. Regular employees can be continued throughout the week. Also, a decision about the number of counters to be open can be taken to reduce operational costs. There should be a smaller number of counters open on Monday and other weekdays as compared to Saturday and Sunday.

  • Monthly forecasts for the 2 products in a particular store will help determine (a) the inventory levels, and (b) the order levels.

As can be seen from the data, sales peak in the month of July and the minimum sales is observed in the months of December and January. A mini peak is also observed in the month of November. Firm should ensure maximum inventory during the month of July and can use its labor force in the previous months for product packaging and systematic updating of SKUs. Firm should be aware of ordering extra products during July and November. To manage extra capacity handling during July, pre-planning about warehouse must be done.

  • Quarterly or annual forecasts will help in capacity building at the aggregate production level.

There has been an increase of 11.06% in the sales 2017 to 2018. Based on the available forecast, the firm should manage its capacity. Firm can also decide whether it can improve its efficiencyand processing to increase troductivity, or it has to add additional capacity.

Application Area: