Estimation of causal dependence between economic openness and deviations from uncovered interest parity using double machine learning

The method of double machine learning was proposed by Chernozhukov et. al. (2018) for estimating structural parameters and treatment effects in statistical models containing a high-dimensional confounding parameter (e.g., control variables with an unknown functional form of dependence). In this work we apply double machine learning for estimation of the relationship between deviations from uncovered parity interest rates and the degree of economic openness, complicated by the nonlinear influence of confounding variables. It is shown that this method, despite weaker assumptions about the data generation process, allows obtaining more accurate estimates, which comply with modern theoretical constructions, and takes into account the heterogeneity of the effect. In particular, our estimate is positive for developing countries, and negative for the group of high income countries - which is consistent with the mechaninsm in Itskhoki, Mukhin (2017), which explains this effect by difference in net export elasticity.