Weather-aware decision making in retail

Applying or not markdowns is a decision of paramount importance that a department stores necessarily have to make in order to keep their loyal customers as well as acquire new ones.

But, when should the supermarket offer discounts to attract customers and when discounts entail a loss of income? Does markdown effectiveness depend on weather conditions? Those are some of the questions that Alex Ceccotti, Davide Pecchia and Michela Sessi, three students of Master of Science program in Data Science at University of Milan–Bicocca, addressed in their final project for the course of Decision Models. In particular, the aim of the study was to create a tree-based decision model to help supermarket store management to decide whether applying markdowns in a certain week depending on predicted weather conditions and other exogenous variables.

A publicly available data set (on the Kaggle web platform) has been selected and used as a playground to test assumptions and develop a suitable decision model. The data set features information about sales, stores, discounts and various features related to external conditions. In particular, weekly data are provided, for the year 2012, of an American store chain featuring different kind of variables ranging from temperature and CPI (Consumer Price Index) to unemployment rate and fuel price.

A decision tree has been developed: the decision to apply or not markdowns for a certain week and a determined store is made by selecting the highest expected value of the actual weekly sales between the two branches of the tree. In order to perform the evaluation, different features have to be predicted. Gross weekly sales are estimated exploiting a regression tree. The most influential variables resulted to be the as the CPI and Unemployment rate, both predicted by two ARIMA models.

The main payoff variable for the realized decision model is the Ratio, that gives an information about the impact of the offers on the sales. It also has to be estimated. This prediction is made by means of another tree that considers variables such as the season, the outside temperature, and the week type according to holidays. Next, a neural network for every last leave of the tree is applied to achieve the goal of improving fitting.

  • In the first branch (related to the “apply markdowns” decision), different results are obtained over three scenarios based on the predicted temperature. The probabilities for the temperature variable are calculated as a Normal distribution, centered in the prediction with a standard deviation of all the temperature of the season.
  • In the second branch (“not apply”) the percentage of discount has been considered: in this data, markdowns correspond to 20% reduction, so in order to get back to the full price, a 25% of increment has been estimated. Finally, another variable has been used: an elasticity parameter. The hypothesis, accredited by various studies, is that by not applying markdowns and, consequently, by increasing the price, there is high probability of loosing customers, thus achieving an overall loss of sales percentage. This risk is estimated using the elasticity parameter that, according to the relevant literature, changes with the types of product, years and other factors known to the domain expert. For this work, a Sensitivity Analysis has been implemented; this tool allows the experts to decide the most suitable approach to the problem.

The model created by these off-the-chart students showed that the temperature has a significant influence on the impact of markdowns on sales. As if this conclusion were not enough, the proposed study also proposes possible extensions that include other weather features (humidity, rainfall, wind speed) to obtain a more accurate model.

Leave a Reply

Your email address will not be published. Required fields are marked *