In the first model we use the 10yr – 2yr spread. Surprisingly, given that the literature finds this spread is useful for forecasting recessions, the statistical relationship between the two variables is insignificant (that is, the correlation is so low that the observed relationship could be due to chance). The fraction of the variance of real GDP growth accounted for by the spread (the 'R-squared') in this model is 2% and thus almost zero.
In Model 2, the 10yr – 3mo spread is used. The results show that it is highly significant, as evidenced by the significance level of 1%. While the R-squared is much higher, 9%, it is still the case that the forecasting power of the spread is limited.
By splitting the 10yr – 3mo spread into the 10yr – 2yr spread and the 2yr – 3mo spread, the information contained in the term structure can be studied more closely. The results, in Model 3, are surprising: all the information about GDP growth over the next four quarters is contained in the 2yr – 3mo segment, which is highly significant. The R-squared also almost doubles.
Non-linearities
The finding that the 10yr – 2yr spread does not forecast future real GDP growth is surprising, given the evidence that the spread forecasts recessions. One possibility is that the relationship is non-linear. It could be that an inverted term structure indicates that future growth with be below average but that a steeply upward-sloping term structure does not signal above average future growth.
Conclusions
What does this model, which summarises the historical behaviour over the last 60 years, predict for the coming four quarters? Looking at current data, the 2yr – 3mo spread indicates a growth rate of 4.1%.6 This is very high – the median projection of the FOMC members is for 2.8% growth in 2022. It should be kept in mind that this is a point forecast and that the margin of uncertainty is high.
