Publication - Forecast combination for euro area inflation: a cure in times of crisis? - by Kirstin Hubrich, Frauke Skudelny
In this paper we compare the forecast performance of different forecast combination methods with that of a range of short-term inflation forecasting models to evaluate in how far the latter can be used to hedge against bad forecast performance in particular episodes, such as the global financial crisis in 2007/2008.
The suite of models that we consider comprises an individual equation framework (an update of Benalal et al. (2004), as also described in ECB (2010)) which is based on individual equations using a number of explanatory variables; a Bayesian Vector Autoregressive (BVAR) model for the five main components of HICP (see Giannone et al. (2014), and also ECB (2010)); Vector Autoregressive (VAR) models of the respective component of the Harmonized Index of Consumer Prices (HICP), with consumer goods PPI, unit labour costs, oil prices, non-oil commodity prices and the nominal effective exchange rate (VAR_C); and a VAR of all 5 HICP components and the aggregate HICP (VAR_U). These models are compared with a range of benchmark models, including an Autoregressive (AR) model, an ARIMA model, a random walk with drift (RWWD) and a random walk as in Atkeson and Ohanian (2001) (RWAO). In addition, we compare them with two univariate models that might capture potential non-linearities, an MA(1) and a STAR model. We use models that provide forecasts of the five main components of euro area HICP on a monthly basis for the period 1990(1) to 2014(6) and most are updated at least on a quarterly basis.
We analyse the results for the full forecast period, but also for the pre-crisis and the postcrisis period and present the evaluation of the aggregation of the component forecasts. We carry out our forecast combination and evaluation based on every third month, as in the quarterly projections of the Eurosystem. We aggregate the disaggregate forecasts to headline HICP inflation forecasts by using the weights that would last have been available and known to the forecaster in real time.
Three different combination approaches are included in our forecast model comparison: the simple average, using equal weights; a performance-based forecast combination using the root mean squared error for a rolling 2-year window of the most recent past to weight the different forecasts; and a performance-based forecast combination as above, with geometrically (backwards) decaying weights, i.e. recent performance is given more weight than performance longer ago. We find that the best model for forecasting differs depending on whether the overall HICP or the HICP excluding food and energy is considered, and which period and forecast horizon is studied.
Therefore we conclude that performance-based forecast combination helps to hedge against bad forecast performance of some of the models in some situations, even though in the presence of large shocks or crises it does not necessarily improve over the best forecast model since the forecast accuracy of all models might, for example, be biased in the same way. Performance-based forecast combination appears to be useful when the models included in the set of models exhibit very different forecast performance over time.
Forecast combination for the full sample period typically improves forecast accuracy over the autoregressive benchmark model for core and headline inflation, and often improves over single multivariate models. The forecast accuracy gain of combinations is largest for inflation excluding energy and food for the full sample. We also find that performance-based forecast combination improves significantly over the simple average for our application. Investigating combination weights and their development over time, we find significant changes in the weights.
Moreover, there appears to be more pronounced evidence for a structural break in the forecast model performance around the time of the recent global financial crisis, in particular for longer horizons. This time-variation in the weights can be seen as capturing non-linearities in the underlying economic relationships. For instance, the importance of certain variables, such as oil prices or labour-market variables, might change over time. In episodes of more volatile inflation, a multivariate model allowing for feedback effects between inflation and its predictors might improve forecast accuracy. The time-variation in the weights assigned to the forecasts from different single models can help to interpret the combination forecast and improve the forecasting models.
Overall, we conclude from our evidence taking into account RMSFE, forecast accuracy tests and turning point predictions that for euro area inflation, first, performance-based combination tends to outperform simple averages, and, second, that performance-based forecast combination protects against bad forecasts from single models, thereby making the forecast more robust.
Directorate General Communications
Sonnemannstrasse 20, 60314 Frankfurt am Main, Germany
Tel.: +49 69 1344 7455, E-mail: media@ecb.europa.eu
Website: www.ecb.europa.eu
page source http://www.ecb.europa.eu
