Use Case Transportation
Web sales forecasting
The project arises from the business need to have an estimation of web sales in advance to anticipate strategic decisions for the company. From there, 2 objectives were set: the first being a monthly sales forecasting model to predict three months ahead, and the second being a daily sales forecasting model to predict 30 days ahead.
Challenges
During the project, various challenges were encountered that had to be addressed:
Drop in sales during COVID
There was a drastic drop in sales during COVID, which poses a significant bias for the time series.
Seasonal Fluctuations
The peaks and valleys caused by Black Friday and Easter in the daily time series
Solution
Before starting to develop the solution, the issue of the COVID period had to be addressed. To do this, a temporal interpolation of the time series was conducted between January 2020 and June 2022. To carry out both the monthly and daily models, an ensemble of basic time series forecasting models was developed.
For daily forecasting, this ensemble works very well for all months without special events. However, during Black Friday and Easter months, it was unable to forecast the peaks and valleys accurately. To try to improve this, a slightly more complex model was used for these months, the lightGBM, which is capable of better forecasting these peaks and valleys.
Tech stack
Results
A monthly web sales forecasting model three months ahead. A daily web sales forecating model 30 days ahead. Business capacity to make strategic decisions with greater advance notice and maneuverability.
3
Months of monthly sales estimation
30
Days of daily sales estimation
