Transforming Sales Forecasting with Adaptive Machine Learning Models and Data Integration

The company focuses on building electrification solutions, data centers, AI enabled industrial IoT solutions as well as intelligent buildings and homes.

THE CHALLENGE

The company followed a lead management process for tracking all sales activities in their Customer Relationship Management (CRM) tool. Each lead was closely tracked from the initial opportunity stage to the Order Booking (OB) and sale conversion stage.

There was an existing forecasting system that predicted the aggregated sales at a region, Business Unit (BU) or activity level. This sales forecasting system was created using 3 Machine Learning (ML) models that had fixed weightages built into them. Only sales data was used for these ML models and there was no link between the existing OB and sales datasets. These results were displayed in a Tableau dashboard which had limited sales Key Performance Indicators (KPIs) and user interactivity. The 3 ML models were not periodically updated and business users believed that the accuracy of this forecasting system was impacted by potential data quality issues.

The company needed a data analytics partner to analyse if there was a direct correlation between their sales data and other existing datasets, which could improve the accuracy of the sales forecasts. They also wanted new ML models to be created that would provide sales forecasts across regions, BUs and activities, at a granular level. The revised solution would require new Tableau dashboards with a comprehensive set of KPIs that would provide detailed insights to their business users.

THE SOLUTION

The Prescience team performed Exploratory Data Analysis (EDA) on the available datasets and found that they could be combined with the sales data for more precise forecasting. During this EDA phase, the team identified several data quality issues, which were highlighted to the company’s Data Management team for resolution.

To achieve the highest accuracy, the team recommended a dual approach which took sales data forecasting and combined data forecasting, with a dynamic weightage mechanism that would provide the final output. As the data linkages between these datasets increased over time, the dynamic weightage would shift from the initial higher score for sales only forecasting, to the combined data forecasting.

The aggregated forecasts were done at a country, region and BU level, while granular forecasts were done at a region, BU and activity level, for a period of 15 months. Different ML models including ARIMA, SARIMAX, XGBoost, Random Forest, Prophet etc were trained to find which one was best suited to each combination of region, BU and activity level.

The team also worked on opportunity prediction, which defined whether an opportunity would be won or lost. The Random Forest and XGBoost models were employed for this.

The solution included a set of Tableau dashboards which provided different sets of users with detailed sales and opportunity level predictions. These comprehensive dashboards included new KPIs that were not available in the existing sales forecasting system.

The different technologies used for this engagement included,

1. Python
2. Tableau
3. Microsoft SQL Server
4. Microsoft SharePoint

THE IMPACT

With the revised sales forecasting system, business users had access to detailed insights on sales data at a granular level, which were previously unavailable to them. While the earlier system only had 3 ML models with fixed weightages, the new system included multiple models that were best aligned with each combination of data requirements. The dynamic weightage model ensured that the accuracy of the overall system would continue to improve on an ongoing basis.

Also, the application was designed in such a way that new source systems could be easily added with minimal code changes. Since the forecasting system provides figures for up to 15 months, the business users can now plan for future budgeting activities.