Weekly sales forecast is a powerful tool on marketing global strategies, enabling the ability to anticipate the need for reinforcing communication and promotions in the following weeks.
Predicting sales accurately becomes significantly more challenging under unstable contexts. Unforeseen events can disrupt retail dynamics and consumer preferences. COVID19 presented unprecedented challenges, and traditional forecasting models proved inadequate in capturing the rapidly changing landscape.
We propose an iterative approach to sales forecasting. As a baseline model, Prophet is based on time series and enriched with business features (e.g. promotional events, weather forecast, store expansion). By incorporating these variables, the baseline can better detect sales patterns and provide initial forecasts.
To generate a final composite forecast, we employ a recurrent neural network architecture known as Long Short-Term Memory model. Considering the errors underlying the sales of the most recent weeks, LSTM leverages its ability to capture long-term dependencies in sequential data.
The result is more accurate sales forecasts that provide valuable insights for promotional management. The baseline model is responsible for detecting sales patterns, while the forecast tuning focuses on adjusting the absolute sales forecast. This is particularly useful in contextual changes.
Developed under a pandemic context, the framework has proved its value in other economic contexts, such as inflation.