Statement
Adding observations from the cross-section of long-term bond yields to the
macro panel substantially sharpens the identification of monetary policy
regime parameters in a regime-switching no-arbitrage term structure model,
even when the structural parameters are theoretically locally identified
from the macro panel alone. Specifically, including long yields reduces the
finite-sample bias of the estimated policy-regime parameters by roughly an
order of magnitude (factor of ~20 in the Bikbov-Chernov simulation study)
and tightens the bootstrap confidence intervals on the Taylor-rule
inflation-response coefficient alpha by a comparable factor.
Evidence summary
The single direct piece of evidence is Bikbov & Chernov’s TSM-vs-SRM
comparison (their Table 1 and the simulation study described in Section 3.3
and the Online Appendix). The TSM is the full term-structure model
estimated on (g_t, pi_t, r_t, y_t^{2Y}, y_t^{5Y}, y_t^{10Y}). The SRM is
the same structural model estimated on (g_t, pi_t, r_t) only — the same
structural parameters, the same regime structure, just dropping the long
yields. Both models are locally identified at their MLE (verified
numerically in the Online Appendix). But the SRM’s bootstrap confidence
interval on alpha(active) runs from 1.50 to 38.0, while the TSM’s
interval is (1.44, 6.90). On simulated data, the bias of the estimated
monetary regime is reduced by a factor of approximately 20 when the long
yields are included.
The intuition is that long yields encode market expectations of future short rates, which depend on the probability that each policy regime is currently active. Adding long yields therefore adds essentially “free” information about the latent regime that is unavailable from the contemporaneous short rate alone.
Conditions and scope
- The result is established for one structural model (NK + forward-looking Taylor rule with regime-switching coefficients + essentially-affine SDF) on one dataset (U.S. quarterly 1970-2008). Extrapolation to other structural models or sample periods is plausible but not directly tested.
- The “factor of 20” figure refers to the BC simulation study and is specific to their parameterization. Other parameterizations would generate different magnitudes.
- The claim is about finite-sample identification quality (bias and variance), not theoretical local identification — both models are locally identified.
- The yield-curve observations must enter the likelihood as bond prices generated by the same SDF that prices the structural shocks. Using yields as exogenous regressors would not produce this identification improvement.
Counter-evidence
- None directly. Sims-Zha (2006) and some related macro-only studies find little or no role for monetary policy regimes, but they do not use a no-arbitrage TSM to discipline the inference and so are not directly comparable. Their negative finding could reflect either weak macro-only identification (consistent with this claim) or genuine absence of policy regime change (would contradict this claim).
- The claim would be tested by replicating the TSM-vs-SRM design in a different macro-finance context. The cre-asset-pricing-model project’s Hamilton-vs-RBPF gap (~1572 nats) is consistent with the same phenomenon (cap rates sharpen identification of latent regimes) but has not been formally bias-quantified yet.
Linked ideas
- bc-style-identification-simulation-cre-cap — BC-style simulation study quantifying the identification gain from adding CRE cap rates to the observation menu
- Future CRE-project follow-up: replicate BC’s TSM-vs-SRM simulation study with cap rates as the analogous “long yield” observation menu and measure the bias reduction in compound-regime identification.
Open questions
- Does the result generalize to compound regime structures with more than two regimes per chain (BC has 2 x 2 x 2 = 8; what about 3 x 3 x 3 = 27)?
- How does the identification gain scale with the maturity range of the yields used? BC use {3M, 2Y, 5Y, 10Y}; would adding 30Y help further or is the marginal information saturated by 10Y?
- Is the “factor of 20” sensitive to the specific Taylor-rule specification, the SDF specification, or the sample length?