Statement
When a Markov-switching rational-expectations New Keynesian macro model is jointly estimated against (i) treasury yields and (ii) commercial real estate cap rates and income growth, the inferred monetary-policy regime probabilities are significantly more persistent and statistically distinguishable from the bond-only specification, and the underlying macro parameters are different at very high significance. In other words, CRE prices contain regime-relevant information that pure term-structure models cannot extract.
Evidence summary
The principal evidence is two within-sample likelihood-ratio tests in the source
paper. The first test rejects the null that the CAP-model macro parameters equal
the TSM-only-model macro parameters (Chi-Squared, 34 df, p < 1e-10) — interpreted
as: forcing the CAP-model macro estimates onto the TSM model is a substantively
worse fit even at the macro-block likelihood level. The second test takes the
CAP model and constrains its macro parameters to the TSM-model values, then
re-estimates only the CRE-block parameters; the constrained CAP likelihood is
again rejected against the unconstrained CAP likelihood at p < 1e-10.
Qualitatively, the smoothed regime posteriors plotted in Figure 3 of the paper
are visibly more persistent in the CAP model — fewer short-lived regime flips,
clearer post-2011 Passive-Rigid classification — than in the TSM-only model.
Conditions and scope
- Within-sample identification, not out-of-sample forecasting. The claim is about how much the joint information sharpens the conditional posterior over the regime, not about forecast accuracy.
- US data, 1992Q1–2014Q2. The result has not been validated on other countries or other CRE indices (Green Street, RCA).
- Two-binary-chain regime structure. The result is conditional on the Bikbov–Chernov / Leather–Sagi specification of the chain. Single-chain alternatives may give different identification gains.
- The discretion chain is not separately identified by CRE. The paper cannot reject equality of the rigid/flexible chain across the two specifications; the identification gain is concentrated in the active/passive chain.
Counter-evidence
- The CAP-model treasury pricing errors (~40 bp) are larger than the TSM-only model errors (~26 bp). This is the standard “fit erodes when more assets are added” pattern (Chiang–Hughen–Sagi 2015) and is not direct counter-evidence to the identification claim, but it does mean that adding CRE is not a free improvement on bond pricing — the gain is in regime identification, not in bond fit.
- The 2005–2010 model–data divergence in CRE cap rates is large and persistent; one alternative interpretation is that the joint model is mis-specified during that window and that the CAP regime posteriors over that period are also less reliable (the paper itself reads the gap as a measurable bubble-like departure from fundamentals, not as model misspecification, but both readings are consistent with the data).
- The paper’s heuristic Sobol-then-local-optimization global maximum is not guaranteed to be the true MLE; the project’s downstream basin-finder experiments confirm that the likelihood is profoundly sloppy and that global recovery from cold starts is not currently achievable on simulated data.
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
Open questions
- Does the identification gain replicate on alternative CRE price series (Green Street, RCA, REIT-based CRE proxies)?
- Does the gain replicate on a richer regime structure with more than two states per chain or a single joint chain?
- How much of the gain is driven by the cross-section of cap rates (Apartment / Industrial / Office) vs. the time-series of any single asset type?