A Vector Autoregression, Machine Learning, Regime-Based Analysis of Swap Spread Determinants

Abstract

This paper seeks to develop a model to predict the movement of interest rate swap spreads on a short-term horizon of one to six months. We construct vector error-correction models that include traditional determinants of swap spread but additionally include and investigate the effects of three new variables: mortgage refinancing as measured by the Mortgage Bankers Assocaition Refinance index, the US Financial Corporate AAA/BBB credit spread, and the US Financial Corporate BBB/3-Month LIBOR spread. We also repeat our analysis but capture the time-varying effect of macroeconomic regimes through the inclusion of two exogenous variables: the Treasury slope and the level of the VIX Index.

We find that a positive shock to the mortgage refinancing applications has a delayed effect, as expected, and ultimately compresses the three-year swap spread by 1 basis point from 6 weeks after the shock to 20 weeks after the shock. In our model without exogenous variables, we found that the MBA Refinance Index was the largest contributor to the three-year swap spread's variance at the long-term horizon at about 8\%. In our out-of-sample one-week-ahead rolling forecasts, this VAR achieved a RMSE of 0.0286.

We also develop a forecasting procedure for VAR models that are estimated with exogenous variables by utilizing an AdaBoost classifier to make next-day predictions of our exogenous variables. The RMSE of our AdaBoost-aided model showed potential to consistently outperform naive forecasts, though further research will need to be done to accurately model how accurate market participants are at making forecasts of exogenous variables.