Title:	ASSESSING MATCHING ERRORS IN PREDICTIVE MODELS
Author(s):	Krzysztof Dzieciolowski and Daniel Marinescu
ISBN:	978-989-8533-66-1
Editors:	Yingcai Xiao and Ajith P. Abraham
Year:	2017
Edition:	Single
Keywords:	Predictive model, matching errors, model performance, confidence level
Type:	Short Paper
First Page:	329
Last Page:	332
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	Acquisition of new customers and selling new products to them are essential conditions for business success. Predictive models, can greatly improve the effectiveness of this task. In a consumer setting, predictive models have been effectively used to acquire new donors for a charity or new subscribers for an insurance company. The success of a binary acquisition model can be defined as an event in which the prospect has become a firm’s customer. To determine such an event, it is necessary to match the external prospects with the firm’s internal customer base. Methods of fuzzy logic, based on the similarity distance between the strings, are often used for matching. However, two errors may occur in matching: false-positive, when the prospect is incorrectly classified as a firm’s customer; and false-negative, when the customer is incorrectly classified as a prospect. These errors introduce bias in the design of the binary model universe. In an extreme case of complete misclassification, the predictive model would predict an event of interest as a nonevent. These errors and their consequences have so far not yet been considered in the modeling research and practice. In the paper, we assess how the classification errors affect performance of the models. The applications of our approach extend to other areas of binary classification such as network security, spam filtering or medical testing. We provide recommendations for model building strategies and illustrate the approach with actual data.
	Go Back