166 – How Can Optimization Make Statistical Modeling More Actionable?
What-If Analysis and Goal-Seek Analysis for Prescriptive Time Series Forecasting
Jane Chu
IBM
Jean Francois Puget
IBM
Many enterprises apply predictive analytics, e.g., transfer function models, Granger causality types of models, to time series data to forecast what might happen in the future. What can they do to turn forecasting into actionable insights? We propose two prescriptive techniques as answers: (1) what-if analysis predicts the possible outcomes based on different choices of actions; (2) goal seek analysis recommends the best course of actions with a desired outcome. Traditional prescriptive analytics can be hard to use as they (i) require users to specify their business problems as optimization models based on the predictive model, (ii) often require additional data. In this paper, we show how both analyses can be done by solving a constrained optimization problem in a system that combines predictive and prescriptive analytics. Moreover, both predictive and prescriptive models can be automatically derived from available history data plus user defined goals for outcome and constraints for actions. This results in a much more consumable form of prescriptive analytics.