Matching expensive supply to uncertain customer demand is a central problem facing retailers, manufacturers, agriculture biotech firms, hotels, and airlines. Too much supply results in unwanted inventory while too little supply results in lost sales and weakens customer retention. Both over- and under-supplying hurts profits, sometimes significantly.

Amr Farahat, a visiting professor at MIT Sloan School of Management, and his colleague Joonkyum Lee, assistant professor at Sogang Business School in South Korea, have developed a new methodology called Approximate Similarity Transformation that firms can use to improve inventory decisions under general, non-parametric models of customer purchase behavior.

Their research paper titled, “The Multi-Product Newsvendor Problem with Customer Choice,” appears in Operations Research, a scholarly peer-reviewed journal.

“What is new is that big data analytics is now providing better predictive models of customer choice behavior than we ever had, models that quantify complex substitution and other behavioral patterns,” says Farahat.

Approximate Similarity Transformation applies a new algorithm that approximates a retailer’s sales forecast with a simpler and tractable upper bound. Professors Farahat and Lee’s research shows that using the bound as a basis for inventory decisions yields better results than working with the original forecast.

The methodology is most suitable for large retailers looking to leverage their data-honed knowledge of consumer preferences.

Farahat says these retailers can implement their methodology fairly quickly using in-house analytics capabilities.

“A retailer stocking thousands of items for the holiday season or an agricultural biotech firm planting thousands of seed variants for the planting season has only one chance to get it right each year,” says Prof. Farahat. “Our research contributes to a set of prescriptive tools that can help retailers and other firms manage supply decisions of large product assortments by combining these sophisticated predictions of customer behavior with new, sophisticated optimization models. Using computational experiments, we have ultimately shown that our proposed methodology generally outperforms existing benchmarks.”

The Approximate Similiarity Transformation code is publicly available here.