Forecasting Economic Downturns in the Scandinavian Countries using The Yield Curve : Exploring statistical relationships and out-of-sample performance using traditional binary response models, and support vector machine models from the field of machine learning

Abstract

Economic downturns (in this thesis defined as recessions and negative output gaps) are costly and it is in the interest of both the Government and private agents to precisely predict them, either for the government to take actions to avoid them, or for private agents to prepare/adjust. Many advanced econometric models are developed for this exact purpose and experts regularly express their opinion about the future of the economy in the financial press. Yet, U.S. data have shown that a very simple probit model taking variables from the yield curve as explanatory variables has successfully predicted recessions in the past. The aim of this thesis is twofold. First, I test whether the established relationship between economic downturns and yield curve variables in the U.S. also hold in the Scandinavian countries. This is done by estimating a static probit model with the yield spread as the independent variable. Second, I test whether estimating a model from the field for machine learning, called Support Vector Machine (SVM), can improve on the forecasts made by the traditional probit models. To be able to compare forecasts I find a probability threshold, W, that produces binary forecasts from the probabilistic forecasts made by the probit models and calculate several performance measures based on pseudo out-of-sample forecasts. The SVM model directly output a binary forecast so the same performance measures can be calculated directly. First, this thesis find that the coefficients for lags of the yield spread (long rate minus short rate) is significant a 5% percent level for all forecasting horizons tested (from one to eighteen months) when recession in Sweden is the dependent variable. For Denmark only shorter time horizons give significant results, while for Norway very few forecasting horizons prove significant. Models estimated with negative output gap, defined using the Hodrick-Prescottfilter, as the dependent variable yields statistically significant coefficients of the spread at almost all lags for Norway and almost non for Sweden and Denmark. I find no evidence that including more than one lag of the spread is useful. Second, the pseudo out-of-sample tests show that economic downturns are better predicted by the yield curve than by lagged returns of a national stock index. This result holds independent of forecasting horizon, lag length and country. I also find that SVM models that take 10 year bond rates and 3 month T-bill rates as input variables in almost all cases outperform the binary forecasts from the probit model. With respect to which country can benefit the most from the models estimated here, the pattern from the statistical tests are repeated, meaning the out-of-sample results are best for Sweden and Denmark when forecasting recessions, and best for Norway when forecasting negative output gaps.