Using Machine Learning to Measure Conservatism

Abstract

Using a neural network, we develop novel measures of conservatism that fits non-linearities and interactions absent in prior literature. The machine-learning measures exhibit (i) fewer economically anomalous observations, (ii) economic associations consistent with existing studies, (iii) less unexplained year-over-year instability, and (iv) higher economic magnitudes consistent with reduced attenuation bias. The measure further reveals intuitive trends toward a secular decline in conservatism in the US. In simulations, linear models perform honorably even in the presence of a complex data-generating process but causal inference based on machine learning is the most robust to misspecification. The approach offers the promise of reducing noise in measurements and designs more powerful tests to assess theories of conservatism.