CMStatistics 2021: Start Registration
View Submission - CFE
A1663
Title: Estimating Bayesian models using simulated data meta-learning Authors:  Sergei Seleznev - Bank of Russia (Russia) [presenting]
Ramis Khabibullin - Bank of Russia (Russia)
Abstract: A simple algorithm is presented for the estimation of Bayesian models that is based on the principles of meta-learning literature. The algorithm consists of two main steps: artificial data generation and fitting a neural network to the variables of interest. In the first step, an artificial dataset is created as a set of samples from the joint distribution of parameters (generated from prior) and data (generated from the data generating process). In the second step, the neural network is trained in a supervised manner to predict variables of interest (such as parameters and/or hidden variables) on the previously generated dataset. It is shown that the algorithm converges to the posterior mean or any other characteristic of posterior distribution depending on the loss function used in the training of the neural network. The main advantage of the proposed method is that trained once it can be used for any dataset without additional training, so the inference of the Bayesian model becomes almost instantaneous. Two examples (stochastic volatility model and new seasonal adjustment procedure) illustrate algorithm properties.