Sales Predictive AlgorithmSales predictive algorithm that enables quick recalculation
Successful cooperation with the polish division of a global manufacturer of car filters. Sales forecasting for 30 thousand sales positions in weekly and monthly periods. The model was trained on data records from the last three years.
Scope of work
This client’s company is a family-owned manufacturer of liquid and air filter systems, intake systems, and cabin filters headquartered in Germany that continues digital and sustainable transformation.
The main challenge in this project was using various machine-learning algorithms like the Holt-Winters model, The Holt Model, and The Medium Model. To achieve desired results we also had to face limitations like short historical data and unobvious and irregular distribution of damsels.
Our main goals on this project were to achieve the same accuracy of prediction as the prediction made by an experienced employee of the company who was fully responsible for the sales plan in the company.
To achieve the goal of the ideal accuracy of prediction, which usually is achievable just by the experienced employee of the company we’ve made the next solution.
During the analytical work, the following models for forecasting time series were considered some types of prediction models: Medium Model, Linear trend Model, Model AR, Model ARIMA, ARIMA Model taking into account seasonality, Exponential smoothing Model, Holt Model, The Holt-Winters Model.
The demanding nature of the data, the fluctuation, and the occurrence of trends was the reason for the prediction to be based on models from the Holt / Holt-Winters family, which take into account these factors while being resistant to anomalies in historical data.
For our client, a german family-owned manufacturer of liquid and air filter systems, we had to create a sales predictive algorithm that enables quick recalculation whenever it's needed and which could achieve the same accuracy of prediction as the prediction would usually be made by an experienced employee of the company.
To make it possible we filtered tens of thousands of data records, as well as updated and deleted data records and time ordering with data splitting into types. During the analytical work, we considered 11 types of prediction models, which took into account factors while being resistant to anomalies in historical data and helped us to achieve the goal of the ideal accuracy of prediction.