On the initialization of adaptive learning algorithms: A review of methods and a new smoothing-based routine

We provide a critical review on the methods previously adopted into the literature of learning and expectations in macroeconomics in order to initialize its underlying learning algorithms either for simulation or empirical purposes. We find that none of these methods is able to pass the sieve of both criteria of coherence to the algorithm long run behavior and of feasibility within the data availability restrictions for macroeconomics. We then propose a smoothing-based initialization routine, and show through simulations that our method meets both those criteria in exchange for a higher computational cost. A simple empirical application is also presented to demonstrate the relevance of initialization for beginning-of-sample inferences.