Case study: Forecast of sales data

For retailers, previewing the daily and weekly sales volumes in the various stores is crucial for conducting purchasing and sales planning. Using structured data along with a forecast model, COSMO CONSULT succeeded in a project that aimed to detect and explain patterns and forecast sales figures.

• Enabling more accurate purchasing and sales planning
• Avoiding out-of-stock situations
• Learning from the past
• Considering external influencing factors

Sales forecasts are among the most common parameters of advanced algorithms. In most cases, it is based on daily or weekly forecasts based on sales, store and special offer data. In order to take into account as many influencing factors as possible, weather data, holiday dates or information about competing locations are used.

For one of the projects, COSMO CONSULT developed a model that forecasts daily and weekly aggregate demand quantities in order to predict store-related sales.

In the retail industry, most of the sales data show direct dependence of the sales figures on the day of the week. Saturday stands out as the strongest selling day, with up to 70 percent of the total weekly sales. As for the weather data, precipitation has a higher influence than temperature fluctuations. The effect is different even for the different types of stores: For example, the effects are sometimes exactly the opposite between a shopping center and stores at greenfield sites. The same applies to public holidays and school holidays.

The first step was to analyze the past data and find out the factors that really matter. For the forecasting, a gradient boosting model (machine learning) was developed in the programming language R. Having a clean database was crucial for the success of the model. Otherwise false patterns would have been found throughout the data. The forecast can either be integrated into the existing infrastructure or used internally for further projects.

Classical sales forecasts are often used as a subfunction of larger solutions. For example, for the dynamic pricing solution of COSMO CONSULT, forecasts of the sales data and price elasticities are first generated in order to carry out a well-founded optimization. Sales forecasts can also lead to important findings for stock planning. They help avoiding delivery bottlenecks or using discounts through early order transactions.