European Central Bank publishes working paper using machine learning to show linear models overstate aggregate shock sensitivity by up to 1.7 percentage points

The European Central Bank published Working Paper No 3107 assessing how firms’ characteristics shape responses to macroeconomic shocks and what this implies for aggregate dynamics. Using a Generalized Random Forest approach on U.S. firm data, the authors find strong nonlinearities and interactions that materially dampen estimated economy-wide responses, implying that standard linear models can overestimate aggregate sensitivity to shocks by up to 1.7 percentage points. The study uses quarterly Compustat data on U.S. listed firms from 1990 to 2019 to estimate sensitivities of sales, investment, debt issuance and market value to business cycle fluctuations and to monetary policy, uncertainty and oil price shocks, conditional on balance-sheet and non-financial characteristics. While average firm sensitivities are broadly similar under machine learning and a linear panel model, the linear model produces a markedly more extreme distribution, with around 50% higher dispersion and 20% higher kurtosis. Aggregating firm responses by economic weight, the paper attributes much of the aggregate difference to the covariance between firm sensitivities and firm size, with larger firms exhibiting lower sensitivities and generating a median 52% reduction in aggregate economic sensitivity; illustrative results include smaller estimated aggregate responses of sales and market value to GDP growth (about 0.3 and 0.2 percentage points lower than linear estimates) and a more muted stock market and investment response to contractionary monetary policy shocks.