Improving the warranty and returns forecast for a leading US electronics manufacturer using batch and real-time data

The problem

Capacity planning is a crucial division in any organization as the repercussions are both monetary and have a direct impact on the brand and trust of the enterprise. With diverse teams, varying data sources in terms of lineage, and the three V’s of big data, the task at hand is often over-simplified and thus not realizing the true potential of data at hand.

Working with the capacity planning team of a leading US electronics manufacturer, we first developed a champion model to forecast the warranty repairs of a laptop model and its associated SKUs, and improved the baseline by introducing real-time streaming data from social media.

The cognitive solution

Traditional approaches are to treat this as a simple textbook Weibull estimation, parameterizing the lifecycle of the product. The capacity planning team and regional centers often have varying forecasts and demand. The problem was developed in three stages – developing a baseline forecast methodology, incorporating real-time social media data into relevant risk measures, and a pseudo-real-time dashboard of weekly warranty forecasts.

1. Baseline forecasting model

• Develop a hierarchical forecast methodology for each state-SKU variant of the product
• Using the 2P Weibull method as an input to the hierarchical forecast model

2. Social listening framework

• Track official social media profiles and develop a risk score based on natural language sentiments on the product in discussion
• Periodically append the risk measures into the data lake as additional metrics

3. Production-consumption

• Role-based consumption of forecasting metrics for regional and state-wise planning teams
• Event-triggered anomalies notified in real-time on seasonal, end-of-cycle surges

The cognitive design

A GCP architecture was developed to collect, analyze, and host the entire framework to be made available.

Using a three-layered modeling approach enables a lower forecast error and ensures an accurate weekly demand. A trend-watching social media listener was deployed to listen to official Twitter and technical blog websites to periodically scrape discussions and customer sentiments. All of the lexical analysis and creation of information groups were developed as microservices for both on-demand usage and reusability across the teams’ use cases.

Key insights

The implemented framework has decreased over-ordering of parts and spares by 3.6% in the first six months of deployment. The planning dashboard now acts as a central planning tool and a consolidated dashboard piloted in two regions. The scale to extend it to the entire national market and international demographics are key initiatives for the customer currently.