Most businesses require accurate revenue forecasts for efficient operations. Reliable forecasts facilitate the operational planning process and enable long-term investments. At Microsoft, we forecast revenue for multiple divisions across lines of business and geographies. To this end, we developed an R package baselineforecast that builds on top the excellent forecast package that deals with univariate time series. nnOur package has four major advantages. First, it computes concurrent forecasts at different horizons without having to refit the models. Second, it admits arbitrary external features external, e.g. macroeconomic data such as oil prices or unemployment rates. Third, it can forecast arbitrary number of series simultaneously. Finally, it employs an ensemble approach in which rolling forecasts from univariate time series methods act as features for a sophisticated machine learning algorithm such as elastic net regression or gradient boosted regression trees. nnWe applied our approach to publicly available data concerning quarterly revenue for the S&P 500 companies. Our primary data source is companies’ Income Statements, which we augmented with macroeconomic time series from the Federal Reserve Economic Data (FRED), such as Real GDP, unemployment levels, U.S. leading indicator, WTI oil price, and a few others. We trained a set of univariate time series models, as well the boosted regression trees using forecasts from the above as features together with macroeconomic data from FRED. In an out-of-sample test the boosted trees regression outperformed the best time series model at longer forecast horizons, while maintaining virtually identical performance for short-term forecasts.
