Amazon says it’s applying AI and machine learning to calculate inflation rates — and to examine the design of randomized trials and experiments. In a pair of academic papers being presented at the 2020 meeting of the American Economic Association in San Diego, scientists at the company detail their work on models that learn correlations between product features and prices and on a blueprint for experiments that require measuring both average effects and spillovers.

The first paper — “New Goods, Productivity and the Measurement of Inflation: Using Machine Learning to Improve Quality Adjustments” — applies AI to the aforementioned inflation rates problem. As Pat Bajari, vice president and chief economist for Amazon’s Core AI group and a coauthor on the two papers, explains, conventional methods of calculating inflation problematically indicate little change in the prices of goods. One solution is hedonic pricing, where the price of a product is factored into several components that can be compared independently.

The researchers’ algorithm factorizes product prices by identifying relationships between their features and prices. If the model were to be trained on data from one year but fed product descriptions from a year later, for instance, it would output the products’ prices according to the earlier valuation.

Bajari says that Amazon could use it internally to analyze business trends, and that third parties (like banks) could apply similar models to products representative of the economy to observe real-time fluctuations. “If you look at a product line, over the course of a year, 80% of the products might vanish,” he added. “When you calculate the rate of inflation, you’re usually doing an annual measure of price changes. But if 80% of products are gone, that measurement can be inaccurate.”

The other paper concerns experimental design. Randomized trials, like those conducted by pharmaceutical companies where some subjects receive drugs and others receive placebos (controls), compare the outcomes of various effects. Unfortunately, they’re liable to spillover effects, in which the tests (e.g., drugs) end up having consequences for the controls (placebos).

In seeking a remedy, Bajari and colleagues devised a framework that identifies spillover effects by ensuring there’s a related control for every test. It describes how to simultaneously measure average effects and identify spillovers within a single experiment, and it generalizes from double randomization while presenting statistical techniques for analyzing the results of such experiments.

Bajari says the approach could be deployed in a range of different contexts, including movie recommendations, ride-sharing services, short-term-property rental sites, home-buying sites, retail sites, and job search sites. At Amazon, it might be used to determine whether demand spikes are affecting product classes as a whole or are limited to certain products, perhaps to test algorithms for calculating how much of a product to restock at a fulfillment center as a function of recent sales rates and supply on hand.

“This type of spillover does not happen in standard medical-drug trials, because one individual taking the new drug does not affect the outcome for another individual taking the placebo. But it is a feature of many experiments at Amazon and similar companies, where we have complex feedback loops,” explained Bajari. “When people do these kinds of experiments, they usually randomize only one variable at a time. We want to go further with this  idea, where we use multiple randomizations to learn supply responses, demand responses, equilibria — all with the goal to keep improving the customer experience.”