From Predictive Modelling to Optimization: The Next Frontier

In “The Evolution of Data Products”, O’Reilly Media’s Mike Loukides notes: “the question of how we take the next step — where data recedes into the background — is surprisingly tough. Do we want products that deliver data? Or do we want products that deliver results based on data? We’re evolving toward the latter, though we’re not there yet.” In this talk, Jeremy Howard will show why taking this step is tough, and will lay out what needs to be done to deliver results based on data. He will particularly draw on his experience in building Optimal Decisions Group, where he developed a new approach to insurance pricing which focused on delivering results (i.e.: determine the optimal price for a customer) instead of delivering data (i.e. calculating a customer’s risk, which had been the standard approach used by actuaries previously).

Delivering results based on data requires 3 steps:

1) Creating predictive models for each component of the system 2) Combining these predictive models into a simulation 3) Using an optimization algorithm to optimize the inputs to the simulation based on the desired outcomes and the system constraints

Unfortunately, many data scientists today are not sufficiently familiar with steps 2) and 3) of this process. Although many data scientists have been developing skills in predictive modelling, simulation and optimization skills are still rare. Jeremy will show how these 3 steps fit together, give examples of their use in real world situations, and will introduce some of the key algorithms and methods that can be used.