Prophet is a flexible Bayesian time series model designed to address the challenge of generating a large set of high-quality forecasts across an organization for a diverse set of business objectives. This challenge is particularly exacerbated when parts of the organization do not have time series experts or when data scientists do not have the requisite domain knowledge.
In the original formulation, these time-based effects are the only regressors in the model, although they may be fit using with nonlinear terms. Additionally, the model includes automatic changepoint detection in for the trend component. Justin Bleich explains how Coatue—a hedge fund that uses data science to drive investment decisions—extends Prophet to include exogenous covariates when generating forecasts, similar in flavor to a classical ARIMAX model, which would include time series terms. Given that the exogenous predictors enter the model as linear terms, interpreting the conditional effects of the predictors on the response is easy for an analyst and the Bayesian framework allows for straightforward uncertainty estimation allows for inference.
Justin then explains how Coatue applies Prophet to nowcasting macroeconomic series using higher-frequency data available from sources such as Google Trends in order to obtain accurate forecasts of headline numbers such as the weekly initial jobless claims number or the monthly retail sales number. These series often exhibit predictable time series effects, such as additional information on consumer and economic behavior like finding the most common search terms on Google Trends that correlate with the macroeconomic series of interest. Given that one of the biggest challenges in macroeconomic forecasting is forecasting changepoints, Prophet may be particularly well suited to the task versus competitor models. Justin showcases Prophet’s forecasting ability when including exogenous predictors and compares it to a number of benchmark models.
Coatue is contributing its fork of Prophet with covariates, dubbed ProphetX, to the open source community. In the future, ProphetX could be modified to include a spike-and-slab prior and stochastic search variable selection to execute automatic variable selection in a Bayesian framework if a large number of exogenous predictors are included.
Justin Bleich is a senior data scientist at Coatue Management. Previously, Justin was the cofounder and CTO of Zodiac, an artificial intelligence startup that focused on predicting customer behavior to help brands retain their best customers and find more like them, and an adjunct professor at the Wharton School at the University of Pennsylvania, where he taught advanced data mining and predictive modeling. Justin holds a PhD in statistics from the Wharton School, where he focused on Bayesian machine learning and ensemble-of-trees algorithms.
©2017, O'Reilly Media, Inc. • (800) 889-8969 or (707) 827-7019 • Monday-Friday 7:30am-5pm PT • All trademarks and registered trademarks appearing on oreilly.com are the property of their respective owners. • email@example.com