We describe an application of predicting sales opportunity using similarity-based methods. Sales representatives interface with customers based on their potential for product sales. Estimates of this potential are made under optimistic conditions and referred to as the opportunity: How much can be sold if a sale were to be made? Since this can never be verified exactly, the direct use of predictive models is difficult. In building systems for estimating sales opportunity, the key issues are: (a) predictions for targets that cannot be verified, (b) explanatory capabilities (c) capability to incorporate external knowledge (d) parallel computation of multiple targets and other efficiencies (e) capability to calibrate optimism in the predictions, (f) method stability and ease of maintenance for incorporating new examples. Empirical experiments demonstrate excellent predictive accuracy while also meeting these objectives. The methods have been embedded in a widely-used similarity-based system for IBM's worldwide sales force.
展开▼
