Huge progress has been made in commodifying both predictive modeling and A/B testing. Less progress has been made in productizing so-called structural models---models based on sufficiently rigorous microfoundations to be able to make informed judgements about the behavior of a system away from the training cases.
A common case of where structural models outperform standard predictive models is in predicting sales for a new product in a competitive market. We might train a model to predict sales based on product characteristics and market conditions. Yet the training data won't typically provide cases of the response that competitors will make in response to the entry of your product.
We must jointly predict the strategic behavior of competitors as well as prospective sales given their strategic response. This necessitates specifying the structural problem that all agents are solving. Jim will provide an example of such model, and talk through the steps required to take these approaches to industry-scale data in a productized way.
Jim Savage is a Manager, Data Science at Schmidt Futures, where he assists portfolio projects in prototyping their data science solutions, and helps to source grant and investment opportunities. Before Schmidt Futures, Jim was Head of Data Science at Lendable, where he built systems to automate due diligence processes and price portfolios of small loans in Sub-Saharan Africa. He also worked at the Grattan Institute in Australia, where he worked primarily on retirement savings policy, and at the Australian Treasury, on the ill-fated 2011 carbon price. He is a Bayesian statistician who specializes in discrete choice, causal inference, time series analysis, and the incorporation of contemporary machine-learning methods into all three fields.