Learning, large deviations and rare events
In: NBER working paper series 16816
"We examine the asymptotic distribution of estimated coefficients and endogenous variables in a dynamic self-referential model when agents learn adaptively using a constant gain stochastic gradient algorithm. The model environment can represent a number of economic models, including asset pricing models, that have been studied recently in the adaptive learning framework. The asymptotic distributions of forecasts and endogenous variables are characterized using techniques from linear recursions with multiplicative noise and large deviations, and are shown to exhibit fat tails"--National Bureau of Economic Research web site