Dominik Hecker

This paper proposes a novel machine learning approach to forecast U.S. inflation. Using signal processing techniques, I engineer informative features from past inflation data to train a Random Forest regression model. This parsimonious approach contrasts with more elaborate machine learning methods that rely on hundreds of time series, and differs from empirical and structural models with dozens of equations traditionally employed by policy institutions. The proposed model outperforms benchmark forecasts from the Survey of Professional Forecasters and Tealbook projections at short horizons in real-time. These findings suggest that a simple, data-driven model would have surpassed complex, large-scale institutional models during the unprecedentedly volatile inflation period after 2020.
Paper

We study countercyclical capital buffer rules that are robust to model uncertainty. We evaluate these rules across 12 Dynamic Stochastic General Equilibrium models, which differ in banking-sector frictions and transmission channels. To ensure comparability, we apply consistent loss functions and policy rule specifications. Our results show that robust rules call for a moderate response to key credit indicators, such as the credit-to-GDP ratio or credit growth. For a one-percentage-point deviation from steady state, the capital buffer should increase by 10-20 basis points. This cautious approach balances restricting credit and curbing growth with preventing excessive credit expansion and financial instability.
Bank of Finland Research Discussion Paper

We estimate a New Keynesian model that allows endogenous transitions between a target equilibrium, with inflation fluctuating around the central bank’s target and interest rates typically positive, and a low-inflation equilibrium, where the effective lower bound binds and de-anchored expectations keep inflation persistently below target. The model is estimated using Bayesian methods, employing an ensemble MCMC sampler with a particle filter to handle nonlinearities. We find that the United States remained in the target equilibrium after the global financial crisis, the euro area transitioned to the low-inflation equilibrium in 2015, with the subsequent inflation surge initiating a return to the target equilibrium in 2021, and Japan entered the low-inflation equilibrium in the early 2000s. Bayes factors strongly favor the equilibrium-transition model over an alternative specification in which the lower bound binds only occasionally and expectations remain anchored.

To quantify the impact of uncertainty on economic dynamics and monetary policymaking in a partial information setting, this paper introduces stochastic volatility and learning into a baseline New Keynesian model. Shock volatilities are subject to uncertainty innovations, which agents gradually learn about. Stochastic volatility and learning amplify welfare losses and bias optimal policy prescriptions. Average Inflation Targeting (AIT) outperforms Inflation Targeting (IT) in mitigating uncertainty and partial information effects. These results underscore the importance of incorporating information frictions in designing effective monetary policies under uncertainty.

Both standard macroeconomic models and professional forecasters systematically overpredict inflation during periods of persistently low inflation. We show that using a model in which inflation expectations can de-anchor from the inflation target, allowing a transition to a zero interest rate and low-inflation equilibrium, would have improved forecasting accuracy in Japan since the late 1990s and the low inflation episode in the euro area between 2013 and 2020. At other times and for the US economy, the model's forecasts remain similar in accuracy to a counterfactual model with anchored inflation expectations and to central bank inflation forecasts.