Simple Bayesian Reference Class Forecasting for Binary and Continuous Business Variables
April 17, 2024
Speaker: Shaun Comfort (Genentech, Inc.)
Various industries generate forecasts to facilitate portfolio investments and decision-making. Typical examples in pharma include binary probability of success estimates for clinical trials and peak sales estimates. The vast majority of these forecasts are generated using subject matter experts and/or quantitative models using detailed information about the project(s) under consideration. From the Heuristics and Biases perspective, these forecasts are based on an “inside view” of projects resulting in predictions that are non-regressive and miscalibrated, relative to actual outcomes. In 1977, Daniel Kahneman and Amos Tversky published a paper outlining a simple corrective procedure for combining inside-view forecasts with distributional data from relevant outcomes, to produce plausible forecasts that are closer to actual results. This approach termed has been termed “reference class forecasting” Flyvbjerg (2011). For this presentation, Dr. Comfort presents a simple Bayesian reformulation of Kahneman and Tversky’s approach to update inside-view forecasts with prior information to produce posterior probability distributions for binary and continuous business variables. He illustrates this with a practical example from his recent article in Foresight Issue #72 (https://forecasters.org/foresight/issues/) to generate posterior estimates of clinical program probability of success and 5-year peak revenues, from initial inside-view forecasts.