Forecasting yield spreads under crisis-induced multiple breakpoints

Abstract

We perform a real time, out-of-sample forecasting exercise concerning seven fixed income spreads sampled at weekly frequency over a sample that spans the US financial crisis. We compare the predictive accuracy obtained from univariate, mean-reverting models of spreads that ignore the evidence of structural breaks in correspondence of the crisis, with models that take estimated and exogenous break dates into account. We also benchmark these predictive performances to standard random walk models. We find little or no evidence that accounts for breaks in the conditional mean process of yield spreads that would have improved real time predictive accuracy. We speculate on the reasons of such failure and we find informal indications that poor estimation of the breakpoint and the higher variance characterizing the post-break period is responsible for our results.

Keywords:

• yield spreads
• structural breaks
• forecasting accuracy
• random walk

JEL Classification:

• E40
• E52
• C23

Notes

¹ For a description of the economic meaning of these spreads, see Guidolin and Tam (2010).

² Even though we have tried initialization windows with more than 9 + 1 observations, 10 was considered to give the minimum number of data points needed to estimate the three parameters in (2).

³ This procedure differs from PT’s (2002) as they suggest using only post-break observations. This is not possible in real time experiments, because a forecaster should stop predicting after each break until the minimum number of observations needed to estimate the coefficients becomes available again. During the 2008–2009 GFT, most FI desks would not have tolerated being left with no trading signals for 2 months.

⁴ ‘(…) in true real time, (…) crazy forecasts would be noticed and adjusted by human intervention’ (Stock and Watson, 1999, p. 4).

⁵ We have also used modified versions of DM that adjust for small sample deviations from normality of the test statistic. However, with roughly 400 forecast errors (2004–2011), this made little difference to our results.

⁶ We considered as outliers all the forecasts that exceeded the maximum/minimum predictive value obtained up to that point in real time. The underlying idea is that ‘(…) a common practice in the literature on large forecast comparisons is to mimic the behavior of a true forecaster by setting an automatic insanity filter (Stock and Watson, 1999). The filter mechanically discards each forecast value exceeding (…) some given threshold and replaces it with some reasonable value’ (Ferrara et al., 2012).

⁷ As summarized in PT (2005), the optimal fraction of pre-break observations to be used in estimation after a break should be higher if the break is small, if the post-break parameters variance is higher than the pre-break one, and if the post-break window is short. So the optimal pre-break window is determined optimizing the trade-off between bias (increasing in the number of pre-break data) and forecast error variance.