Optimizing Product Offers with Machine Learning and the Mixed Logit Model
By John Colias and Beth Horn


How does one create an efficient analytical pipeline that begins with data for responses (sale or no sale) to real offers made to customers and ends with an optimal offer of product and price to any current or potential customer visiting a retail site? This presentation will outline an analytical pipeline that uses machine learning to optimize a mixed logit model and then applies nonlinear programming to optimize the offer. The advantage of the mixed logit model vs. other popular machine learning models is that the mixed logit model is grounded in economic theory and thus produces sound pricing recommendations. The predictive accuracy of the mixed logit model will be compared to that of other types of models used for machine learning. Data requirements, estimation and optimization methods will be presented with empirical evidence using transactions data.

Contact Decision Analyst

This presentation was made to the 2021 informs Business Analytics Conference by John Colias, Affiliate Assistant Professor of Business Analytics at the University of Dallas & Senior Vice President at Decision Analyst and Beth Horn, Senior Vice President, Advanced Analytics at Decision Analyst.

They may be reached at 1-817-640-6166.



Video Chapters

Below are the timestops for the video:

  • 00:35 Our Story for Today
  • 02:09 Mixed Logit and Random Forest Models
  • 09:48 Comparison with Random Forest
  • 11:42 Case Study: E-Commerce Data
  • 13:56 Modeling The Results
  • 24:01 Optimization of Price
  • 28:00 Summary of Analytical Pipeline
  • 28:43 Case Study Conclusions


  • John Colias
    John Colias
    Senior Vice President, Decision Analyst
    & Affiliate Assistant Professor of Business Analytics at the University of Dallas jcolias@decisionanalyst.com