Utilising weather in demand forecasting

Problem: From October 2011 through to August 2013 we provided the Waitrose supermarket chain with advice on how to incorporate weather forecasts into their demand forecasting system. Weather is known to have a significant impact on demand and can lead to increases or decreases in demand across a wide assortment of lines. Sales of salads and burgers clearly increase when the weather is good for barbeques, whilst sales of soup increase when the temperature drops. Other less obvious lines impacted by weather include cat litter, which has proved effective in gritting driveways during icy spells. Utilising weather forecasts helps ensure the right lines are stocked at the right time, thus reducing stock-outs as well as reducing wastage.

Approach: Our work was undertaken in four distinct phases:

  • Assessing the options available for utilising weather forecasts;
  • Undertaking a feasibility analysis of the preferred option;
  • Designing a statistical model and producing documentation for implementation; and
  • Providing technical advice during implementation.

Out initial role was to advise on the best solution to pursue. This involved examining the current processes, liaising with stakeholders, researching the approaches used by other food retailers, and meeting with potential software vendors and potential weather forecast providers. The resulting recommendation was for Waitrose to develop an in-house solution with the aid of our expertise.

During the feasibility phase we obtained and formatted the large volumes of raw data subsequently used to populate our statistical model. By identifying correlations between historic weather data and historic sales we were able to specify a model which adjusts the system generated demand forecasts according to a set of forecasted weather parameters, including maximum temperature, sunshine hours, rainfall and snow. We demonstrated a final prototype model to senior decision-makers and presented them with expected cost savings. We were then asked to proceed with full model specification.

Methodology: We developed a detailed solution which operates at the line-branch level in order to take into account the characteristics of each line at each branch location. We wanted to make the most of what was sometimes sparse data, so we also used clustering techniques to combine branches and lines which share similar demand responses to weather. Such information is useful when new stores are opened or new lines are introduced, and historical demand data would otherwise be lacking.

We designed a number of performance metrics to measure the benefits that the improved demand forecasts would bring to the business. These made use of traditional forecast accuracy measures, including the commonly utilised mean absolute percentage error (MAPE), alongside the mean absolute scaled error (MASE) which was obtained from recent academic literature and better takes into account the demand patterns we observed. We also developed a simulation model to mimic the stock ordering system under a new forecasting regime, and measured the expected quantity and value of lost sales due to stock-outs on the one hand and wastage due to unsold products going beyond their sell by dates on the other.

Outcome: Following our work Waitrose have undertaken a major weather implementation project to replace the existing manual weather approach which can only consider the impact on a limited range of products. Our design for a new system offers a sophisticated and automated method for incorporating daily weather forecasts up to 14 days ahead. We worked as implementation advisors during this phase to ensure the methodology was correctly specified and to ensure any unforeseen issues could be resolved quickly.

Our analysis showed that, for those lines where sales were affected by weather, lost sales could be reduced by 6% and wastage by 1%, leading to a reduction in costs of around 2%. As the supermarket business is a low-margin industry where the average profit margin typically ranges between 1 and 2 percent, this result was considered significant.

Client comment: The client commented, “Andrew was fundamental in designing the right forecasting solution to manage seasonality and weather.  Critically, as well as the solution design he helped show the forecasting estimated benefits which was essential to enable key stakeholders to make the decision to invest in the work.  Monthly seasonal multipliers have gone in successfully with forecast improvements logged.  We are now heading for the trial of the Weather element.” – Gail Milner, Waitrose Supply Chain Portfolio Manager, May 2014.

Benefits: Our work with Waitrose has allowed them to implement a weather model which leads the food retail industry, and the experience we have gained can be applied elsewhere. The work involved communicating with client staff at all levels in order to develop a methodology that would be acceptable to everyone. We had to present clear and convincing results in order to obtain Board level approval to progress through the various phases. We found that collaboration worked best and that actively sharing our knowledge enabled the client to learn and contribute more by being closely involved. Furthermore, ensuring everyone understood the methodology and assumptions, and providing clear documentation, prevented any doubts arising from the client side.

Contact: Andrew Eaves