How weather works – or doesn’t – for brick-and-mortars

Sometimes weather doesn’t mean a thing, as for example in the grocery segment. Quite the opposite goes for some other businesses, for example retailers in the segment of Consumer Electronics & Home Appliances – it doesn’t take data to conclude that bad weather to some extent impacts foot traffic behavior more than good weather and (might) mean stronger sales. Not only weather, there are also a number of other “unusual” factors that can all impact customer demand and should be noted when trying to anticipate customer traffic. Is there for example a school holiday or some other shopping event that is going to happen in the coming time? Foot traffic is also a major metric for brick-and-mortar campaign success and can be mined for determining what’s working and what isn’t.

Big Bang Ltd., the biggest Slovenian Consumer Electronics Specialist has also noticed the impact of previously mentioned factors. For example, prolonged winter into the month of March 2018 increased foot traffic in stores for 10% compared to the same period PY as well as sales in all categories. Another example is Black Friday shopping event that transferred Christmas shopping fever to November. In 2018 Big Bang has experienced 17% greater foot traffic store compared with PY and 69% stronger figures compared to average foot traffic during all Fridays in the same month.

For businesses such as Big Bang, being able to understand the connection between good and bad weather, different events and brick-and-mortar visitation, also provides some valuable insights. The relation between visitor numbers and profits has been researched extensively and shown to be vital. This results in a need for retail businesses to predict visitor numbers accurately and sufficiently in advance. If such businesses can better anticipate foot traffic flows as a result, they can stay ahead of their competitors and make it rain in the positive sense. Weather can provide predictive clues to what consumers might do and add an additional point of differentiation from its competitors. Also, if retailers spend time identifying variables that impact their customer traffic before starting to build shift schedule, with the right scheduling tools they can create the right sales forecast, schedule staff to match this customer demand and improve customer service or at least make it consistent.

Therefore, the main business opportunity lies in the lack of systemized evaluation on how weather and events influence daily business and the need for Weather and Event-aware Business Intelligence to optimize business decisions. The knowledge of predictive modelling can be applied to optimize resources and ultimately help to maximize marketing budgets and optimize workforce. Based on model results Big Bang aims for optimization of Salesforce on weekly level i.e. extend weekends off-work for retail staff compensated with stronger sales force presence, depended on the weather forecast.

The services will also determine how to manage activities in given conditions i.e. weather and external events – when to enforce “must have” activities vs. »best performing« activities and when are “peak” weekends for promotional activities to be offered to distributors and “friendly” brands.

Predicting the unpredictable consumer behavior

But how exactly can we add predictability to unpredictable consumer behavior, which shifts as fast as the weather does (and so should marketing and resourcing)?  What timeline and geolocation depth of data needs to be examined?

A robust prediction system is needed to cope with various foot traffic patterns dependent on mixture of weather and event variables. It should also be effective in modelling weekdays and weekend patterns, seasonality effects and capable of accurately predicting number up to a few days in advance to enable weekly planning.

Since Big Bang has stores on multiple locations with different types of premises (stand-alone vs. located in malls), model should also be adapted to variations across (micro)location with transferring information from “bigger” to “smaller” stores (where size can refer to foot traffic and/or number of employees). We should be able to answer whether our best-performing store is generating more sales simply because it is in a location that receives higher traffic?

A further step is also needed to study the relationship between store traffic and sales performance, which should be decomposed in sales value and volume, closing rate (defined as the ratio of number of transactions to traffic) and basket value (defined as the ratio of sales volume to number of transactions) and analyze the impact of traffic on sales and its components. As can be seen from the Big Bang data a high correlation can be noted there.

How weather affects sell-out

In the Pilot Phase the project was set with focus on predicting how weather affects sell-out on category and SKU level. Within the Pilot an integrated platform (based on MariaDB technology) was built, where data from BC partners Ceneje and Big Bang were merged on daily basis (with 24 hours delay). Data were further enriched via API with weather data sets to predict how weather affects customer interest (measured through pageviews on Ceneje website & deeplinks to Big Bang website) and sell-out (for Big Bang offline and online stores merged). We ran 28 different predictive models (offset for 7 days) in real time to be able to daily monitor, cross-check and verify the output. Datasets included data for two chosen categories (TV and AC), for all the products that are active and listed on Ceneje website. Predictions were firstly run on SKU and then also aggregated to category level.

Further on, we have monitored the precision of predictions through different measures and evaluated the output from business perspective. The Pilot results have shown that the prediction of consumer interest and sell-out on category and SKU level needs focus shift to a higher level i.e. channel/category demand level. Algorithms were also too much focused-on weather, which is important but not always the most influential factor and is also category specific . Therefore, in the next steps the existing models needed to be improved firstly by including weather demand on more aggregated business features like traffic, conversion rate and basket value and secondly by enrichment of predictive models with internal and external events.

In the second Phase of the project Linear Regression modelling was employed in-house. Foot traffic predictions were run on the historical source data for five years on a daily level (from 2014-2018). In the datasets we have utilized historical record of consumer counts captured with sensors and decomposed sales data, which were put in the model as independent variables. To identify variables that (potentially) impact customer traffic and determine retail connected data points, external conditions were considered.

Datasets were enriched with time series of daily values for different external events (categorized to Calendar Holidays, School Holidays, Shopping Events and Sports Events) and internal marketing events (grouped by channel) and linked to weather data, which were geolocalized (for exact locations where we have stores) or with one central weather station (national weather).

Additionally, 30 years average of national weather data was given in the model, to enable calculations of weather condition deviations. We have prepared different datasets and data transformations, e.g. offsetting events, following seasonal differences etc. Multiple Linear Regression (Forward) was employed for traffic predictions and showed some promising results and interesting points that need to be further examined. The data that was added to the dataset turned out to have strong correlation and strong periodicity per season, (week) day and (micro)location. There were no differences whether weather was taken geolocalized or for central weather station and better accuracy was achieved when weather variables were more „straightforward“.

Since the accuracy of these models leaves room for improvement and the complexity of all variables included can’t be addressed with linear modelling in SPSS, full AI modelling was provided by IJS on training and testing sets with weather information included or excluded.

Applying weather forecast to predict foot traffic

The use of meteorological and event data should allow our business to have a clearer response to shifting foot traffic patterns with objective of improving Campaign Management and Salesforce Management. Therefore, the solutions are mostly aimed at two different internal departments i.e. sales and marketing department.

Applying weather forecast to predict foot traffic can result in at least four improvements that Big Bang will experience:

  • more accurate sales realization forecasting;
  • adapting marketing activities in such way to make ad-hoc activities more beneficial (i.e. applying activities aimed at increasing volume when we anticipate higher foot traffic or activities aimed at increasing basket value when foot traffic is expected to be lower);
  • workforce schedules can be adjusted more precisely to improve profitability (which also satisfies the legal obligation of changing schedules three days in advance);
  • work on the floor can be more efficiently organized by adapting delivery times when lower visitation is anticipated which furthermore positively affects on realization during high visit times, since goods are already exposed on the shelves.

Currently all the activities mentioned above can only be adapted to PY trends. By connecting historical data with weather forecasting more accurate foot traffic predictions would result in higher realization, better working climate, improved work organization on the floor and coping with logistics and stock challenges more effectively.

Key KPI’s used to monitor the performance are improved closing rate (due to salesforce optimization) and increased sales value (due to marketing activities optimization). Both measures should be improved by two percentage(points) by end of fourth quarter in 2019 or three months after implementation. The existing targets should later be extended based on internal performance track.

 

Interest in this solutions? Get in touch with us!

Patricija Filipič Orel: Patricija.Orel@bigbang.si 

One response to “How weather works – or doesn’t – for brick-and-mortars”

Leave a Reply

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