Yield Curves and Macro Variables Interactions and Predictions

Abstract

This research is based on the yield curves and five macro variables, namely equity indices, FX rates, central banks’ policy rates, inflation rates and the GDP growth rates, for nine different markets, from different geographical regions. Our aim was to identify common trends in yield curves and macro variables behaviors, from two perspectives: the interaction and predictive power of the variables. Firstly, we studied the interaction between yield curves and macro variables based on: Granger Causality, Impulse Response Function and Variance Decomposition. Afterwards, we predicted yield curves based on ANN Regression Multitask learning, and lastly, we predicted our five macro variables based on three different ANN Classifiers, in order to generalize and present results that are not specific to a country, or region, or model. The most persistence trend, amongst the variables, was the association between the GDP, inflation, policy rate and the Level. Based on Multitask learning, we achieved a 1-mnth average yield curves prediction accuracy of 80.2% for all yield maturities and studied markets. Additionally, we found out that increasing the hidden nodes led to overfitting the data, hence, we recommend the use of a simple neural network architecture. Furthermore, we designed a model that computes the optimum number of hidden nodes based on: the number of input/output nodes and forecasted months ahead. The Independent Variable Contribution analysis increased the weight of Slope on average for all markets. Weighted KNN caused a deterioration in the prediction accuracy of macro variables, and K of KNN increased with the horizon forecasted. In terms of predictive power of the variables, the yield curve on its own had predictive powers over long term equity markets, and the policy rate seemed to be affected by macro variables in the short term. Furthermore, the inflation and GDP were dominated by their own past values.