Forecasting using targeted diffusion indexes

Abstract

The simplicity of the standard diffusion index model of Stock and Watson has certainly contributed to its success among practitioners, resulting in a growing body of literature on factor-augmented forecasts. However, as pointed out by Bai and Ng, the ranked factors considered in the forecasting equation depend neither on the variable to be forecast nor on the forecasting horizon. We propose a refinement of the standard approach that retains the computational simplicity while coping with this limitation. Our approach consists of generating a weighted average of all the principal components, the weights depending both on the eigenvalues of the sample correlation matrix and on the covariance between the estimated factor and the targeted variable at the relevant horizon. This 'targeted diffusion index' approach is applied to US data and the results show that it outperforms considerably the standard approach in forecasting several major macroeconomic series. Moreover, the improvement is more significant in the final part of the forecasting evaluation period. Copyright © 2009 John Wiley & Sons, Ltd.

Forecasting Portuguese GDP with factor models: Pre- and post-crisis evidence

“…This modeling strategy avoids discarding potentially relevant information contained in the dataset and tries to obtain a better match between the available data and the variable to be forecasted. As shown in Dias et al (2010), this approach proved to be quite promising vis-à-vis the diffusion index model, improving considerably the forecast performance for several US macroeconomic variables.…”

“…This can result in an important shortcoming for forecasting purposes as such an approach does not take into account neither the specific variable to be forecasted nor the forecast horizon. This shortfall was circumvented in Dias et al (2010) where the authors propose a targeted diffusion index (TDI), which reconciles both the spirit of the Stock and Watson approach and the targeting principle discussed by Bai and Ng (2008). Basically, the suggested procedure considers in the forecasting model a synthetic regressor which is computed as a linear combination of all the factors of the dataset, that is…”

“…However, since it ignores entirely the information content other than the one conveyed by this small set of factors, it may potentially disregard data that can be useful for the variable to be forecasted or the forecast horizon under consideration. Dias et al (2010) suggest an alternative procedure to overcome the above mentioned shortfall. In particular, a tailor made targeted diffusion index (TDI) dependent on the variable to be forecasted and the forecast horizon is proposed.…”

Forecasting aggregate retail sales: The case of South Africa

“…Nevertheless, the current paper adopts an univariate approach, with retail sales being predicted by its past values only, either in a linear or nonlinear fashion. However, as indicated by Dias et al (2010), retail sales are likely to be affected by large number of economic variables. Hence, future research would aim at using linear and nonlinear models of forecasting retail sales involving macroeconomic and financial variables as possible predictors, and, in turn comparing the results with the various univariate models discussed in this paper.…”

Forecasting the Estonian rate of inflation using factor models

“…Periods of stronger co-movement can be expected to yield better forecast performance (Eickmeier & Ziegler 2008). Dias, Pinheiro, & Rua (2010) point out that including only the first few factors in the forecasting equation might exclude other factors that have a high correlation with the target variable or the forecast horizon.…”