TY - JOUR

T1 - Hyperparameter estimation in forecast models

AU - Lopes, Hedibert Freitas

AU - Moreira, Ajax R.Bello

AU - Schmidt, Alexandra Mello

N1 - Funding Information:
We are grateful to Stanley Azen (Editor), Peter Müller, Jonathan Stroud and an anonymous referee for comments and suggestions that considerably improved the article. The first author was partially supported by grants from IPEA and CAPES, Brazil.

PY - 1999/2/28

Y1 - 1999/2/28

N2 - A large number of non-linear time series models can be more easily analyzed using traditional linear methods by considering explicitly the difference between parameters of interest, or just parameters, and hyperparameters. One example is the class of conditionally Gaussian dynamic linear models. Bayesian vector autoregressive models and non-linear transfer function models are also important examples in the literature. Until recently, a two-step procedure was broadly used to estimate such models. In the first step maximum likelihood estimation was used to find the best value of the hyperparameter, which turned to be used in the second step where a conditionally linear model was fitted. The main drawback of such an algorithm is that it does not take into account any kind of uncertainty that might have been brought, and usually was, to the modeling at the first step. In other words and more practically speaking, the variances of the parameters are underestimated. Another problem, more philosophical, is the violation of the likelihood principle by using the sample information twice. In this paper we apply sampling importance resampling (SIR) techniques () to obtain a numerical approximation to the full posterior distribution of the hyperparameters. Then, instead of conditioning in a particular value of that distribution we integrate the hyperparameters out in order to obtain the marginal posterior distributions of the parameters. We used SIR to model a set of Brazilian macroeconomic time-series in three different, but important, contexts. We also compare the forecast performance of our approach with traditional ones.

AB - A large number of non-linear time series models can be more easily analyzed using traditional linear methods by considering explicitly the difference between parameters of interest, or just parameters, and hyperparameters. One example is the class of conditionally Gaussian dynamic linear models. Bayesian vector autoregressive models and non-linear transfer function models are also important examples in the literature. Until recently, a two-step procedure was broadly used to estimate such models. In the first step maximum likelihood estimation was used to find the best value of the hyperparameter, which turned to be used in the second step where a conditionally linear model was fitted. The main drawback of such an algorithm is that it does not take into account any kind of uncertainty that might have been brought, and usually was, to the modeling at the first step. In other words and more practically speaking, the variances of the parameters are underestimated. Another problem, more philosophical, is the violation of the likelihood principle by using the sample information twice. In this paper we apply sampling importance resampling (SIR) techniques () to obtain a numerical approximation to the full posterior distribution of the hyperparameters. Then, instead of conditioning in a particular value of that distribution we integrate the hyperparameters out in order to obtain the marginal posterior distributions of the parameters. We used SIR to model a set of Brazilian macroeconomic time-series in three different, but important, contexts. We also compare the forecast performance of our approach with traditional ones.

KW - Bayes factor

KW - Dynamic modeling

KW - Hyperparameter

KW - Litterman's prior

KW - Posterior distribution

KW - Sampling importance resampling (SIR)

UR - http://www.scopus.com/inward/record.url?scp=0033075220&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=0033075220&partnerID=8YFLogxK

U2 - 10.1016/S0167-9473(98)00078-4

DO - 10.1016/S0167-9473(98)00078-4

M3 - Article

AN - SCOPUS:0033075220

VL - 29

SP - 387

EP - 410

JO - Computational Statistics and Data Analysis

JF - Computational Statistics and Data Analysis

SN - 0167-9473

IS - 4

ER -