Two proof-of-concept projects were undertaken to demonstrate the effectiveness of a new forecasting.
The initial proof-of-concept was to take a sample of 4 months of historical Sales data across 20 stores and see if we could build a forecasting algorithm which improves on their existing one.
Over a few weeks of trial and error, testing different ML software and a myriad of variables we managed to come up with a new algorithm incorporated a multitude of parameters, including day, week of the month, month, payday weekend, bank holiday, promotion, and weather. Furthermore, it employed time-series forecasting, allowing it to consider the previous month’s trend when forecasting for the current period. This enabled the use of all Historical data to train the model, and not be hindered by Sales patterns for Items and Stores changing year upon year.
We eventually settled on software called ‘A Bot Named SUE’ (ABNS), which is perfect for the implementation we chose, and perfect for the challenge facing our client. This software is unlike most other ML software as it is suited and optimised for small data. This allowed us to start predicting with as little as 14 data points, so 2 weeks after a new Item hit the shelves, we could start producing forecasts for it. It was also extremely quick and lightweight, which would enable us to be able to produce the number of forecasts within the timeframe available each day.
Following the initial proof-of-concept, they were happy with the results, but wanted to know how it would perform during, historically, the most difficult periods to forecast (i.e., Christmas), so we performed a similar process and built a model that produced accurate results for seasonal periods (models that were unique to Christmas, that included different parameters such as ‘Days to Christmas’ etc).
Happy with these results, the Client then gave the green light to go with a full build and rollout which would lead up to 500 Item ranges and 1000 Stores nationwide.
A distinctive feature of the solution was its full configurability with a user-friendly front-end application that enabled real-time analytics, error logging, model health/diagnostics, and full configurability. The process, while being completely automated, allowed granular adjustments such as the length of the forecast period, the length of the Trend, the days that go into training the model (extreme periods can be removed that may prove to be unreliable predictions for the future such as certain times during COVID). Providing the Retail company with maximum control and flexibility.