Structural break, biased estimation, and forecast error: Canadian vs U.S. GDP

In a joint attempt to obtain a high level of comparability in collecting business statistics among the Canada, Mexico and the USA, the U.S. Department of Commerce Bureau of Economic Analysis (BEA) switched reporting gross domestic product (GDP) and other national accounts from the Standard Industrial Classification (SIC) System to North American Industry Classification System (NAICS) in 1997. Luitel and Mahar (2015) show that this has resulted in a structural break in U.S. GDP in 1997. A structural break implies that parameter values governing the data generating process have changed, which has important implications in macroeconomics modelling. For example, when researchers employ GDP time series in their macroeconomic models, the unidentified structural break may result in previously unsuspected problems. Luitel and Mahar (2015) have addressed several issues concerning this structural break. However, a vital question still remains: What would be the forecast errors if a researcher did not include the 1997 structural break in the analysis? This study attempts to address the question by examining the same data set of Luitel and Mahar (2015).

Moreover, it is possible that unidentified in-sample breaks may also result in biased estimates of parameters adversely affecting the model’s performance for out-of-sample forecasting. To investigate this possibility, this study also examines the possible poor performance of a forecast and biased estimation in the presence of 1997 structural breaks in U.S. GDP. We gathered this data from the U.S. Department of Commerce, Bureau of Economic Analysis. For this, we first use other stability diagnostic criteria to confirm the structural break in the data. We then fit prediction models with and without break to explore any possible errors in the specification and in the forecast. In order to confirm a break in the structure of U.S. GDP (1973 to 2014), we ran the Luitel and Mahar (2015) model and tested the 1997 break point using a Log Likelihood ratio and Wald statistics. Both statistics indicated that the coefficients were not stable across times, therefore the null hypothesis of no break at the 1997 break point could not be accepted. Finally, we visualize the results. Overall our exercises suggest that predictions from the break models are superior to models that ignore breaks.