Is Accounting Useful for Forecasting GDP Growth? A Machine Learning Perspective

Abstract

We provide a comprehensive examination of whether, to what extent, and which accounting variables are useful for improving the predictive accuracy of GDP growth forecasts. We leverage statistical models that accommodate a broad set of (341) variables---outnumbering the total time-series observations---and apply machine learning techniques to train, validate, and test the prediction models. For near-term (current and next-quarter) GDP growth, accounting does not improve the out-of-sample accuracy of predictions because the professional forecasters' predictions are relatively efficient. Accounting's predictive usefulness increases for more distant-term (three- and four-quarters-ahead) GDP growth forecasts: they contribute more to the model's predictions; moreover, their inclusion increases the model's out-of-sample predictive accuracy by 13 to 46%. Overall, four categories of accounting variables---relating to profits, accrual estimates (e.g., loan loss provisions or write-offs), capital raises or distributions, and capital allocation decisions (e.g., investments)---are most informative of the longer-term outlook of the economy.

Keywords: Accounting; Big Data; Elastic Net; GDP Growth; Machine Learning; Macro Forecasting; Short Fat Data

JEL Classification: E01, E32, E37, E60, M41