higher in the threshold model than in the single equation model; in addition, R2 is also higher in
each sub-regime than in the single regime.
The advantages of examining the EH in a multiple regime framework are evident. The estimated
slope coefficient gets significantly closer to one in both regimes when the threshold variable is the
term premium. When the absolute values of the term premium discriminate regimes, at short and
medium horizons (n = 36, 24, 12, 6) the slope coefficient is statistically significant only in the
regime characterized by moderate uncertainty; while, regime 2 estimates of the slope turn out to be
not significant. Evidence thus highlights a clear asymmetric effect in the empirical analysis of the
expectations theory.
These results can also be interpreted consistently with the hypothesis that interest rate
unpredictability affects the empirical corroboration of the EH (Mankiw and Miron, 1986).
Separating regimes on the basis of the term premium allows investors to identify two distinct states
of the world, each characterized by a specific level of uncertainty; in particular, in both states the
range of values assumed by the term premium is bounded and term premia volatility limited (Table
8). Hence, the investors’ forecast ability improves in both regimes, since in every regime the
variability of term premia is lower than in the single regime. This is true only when regimes are
determined by the level of term premia though (central column of Table 6). However, this no longer
holds when the absolute level of term premia drives the regime switching; as pointed out before, for
medium-short maturities only below the estimated threshold the slope coefficient is significant, i.e.
when prediction errors (tptn,m) are low in absolute value.
In Figure 7 we plot the estimated slope coefficients β against maturity together with the 95%
confidence interval bands. In left part of Figure 7 we report the linear model β estimates; while in
the second and third columns there are the plots of the slope coefficients in the first and second
regime respectively. The second and the third diagrams in the top panel report the regime one and
two β estimates obtained in the threshold model when the term premium is the threshold variable.
In the bottom part of Figure 7 the second and the third diagrams plots regime one and two β
estimates obtained when the threshold variable is the absolute value of the term premium. In the
single regime model at the very short end the yield spread does not have any predictive power21;
while in the two-regime framework the term structure appears to be informative about future
21 Rudebusch (1995) documents significant predictive power of the spread at very short horizons, i.e. lower than two
months. Our analysis is not comparable with his study since we do not consider maturities shorter than 3 months;
moreover, data used in this paper have different frequency.
20