Problem
Standard one- and two-factor affine (CIR) term-structure models systematically fail along three dimensions simultaneously: (i) they cannot match the observed conditional volatility of yields, (ii) they cannot reproduce the observed time-varying correlations between yields of different maturities, and (iii) they reject the expectations hypothesis (EH) in the wrong way (Campbell-Shiller regression slopes are negative and large in absolute value, while affine CIR predicts mild positive deviations). The paper asks whether a parsimonious extension that allows the parameters of an otherwise-standard CIR model to switch between a small number of latent regimes — interpretable as business-cycle states — can resolve these three failures jointly inside a no-arbitrage pricing framework.
Key idea
Embed a 2-state Markov regime-switching mechanism inside a square-root (CIR-style) term-structure model so that the drift, volatility, and market price of risk parameters are regime-dependent, while the no-arbitrage bond pricing equation is preserved within each regime. Bond prices remain (approximately) exponential-affine in the short rate conditional on the regime path, with regime-shift premia entering through the no-arbitrage restrictions across regimes. The latent regime is identified jointly with the structural parameters by the Efficient Method of Moments (EMM), in which an SNP auxiliary score generator distills the data into a moment vector that the structural model is then calibrated against. The estimated regimes line up with NBER business-cycle phases.
Method
- Model class: continuous-time square-root short-rate process with drift, mean-reversion, volatility, and market-price-of-risk parameters that are functions of a 2-state Markov regime variable s_t ∈ {1, 2}. Regime transitions follow a discrete-time Markov chain (constant transition matrix).
- Bond pricing: within each regime, the CIR ODE for log bond prices admits a tractable solution; across regimes, no-arbitrage requires the regime-shift price of risk to enter the cross-regime restrictions. Closed-form bond prices are not available globally; the paper uses a log-linear (affine) approximation in the short rate, which is shown to be accurate for the maturity range of interest.
- Estimation: Efficient Method of Moments (Gallant-Tauchen). Step 1 fits a flexible SNP (semi-nonparametric) auxiliary model to the observed yield panel; the SNP scores form the moment conditions. Step 2 simulates from the structural regime-switching CIR model and chooses structural parameters to make the simulated SNP scores as close as possible to zero in the EMM metric.
- Identification of regimes: filtered (and smoothed) regime probabilities are constructed from the estimated structural parameters and compared against NBER recession dates.
- Tests: Campbell-Shiller-style EH regressions are run on data simulated from the estimated regime-switching model and compared against the same regressions on US Treasury yields, to check whether the model reproduces the negative EH slopes empirically observed.
Results
- The 2-regime CIR model matches the level, slope, and curvature of the US Treasury yield curve and the unconditional first and second moments of yields better than single-regime CIR benchmarks.
- It captures the time-varying conditional volatility of yields that single-regime CIR cannot reproduce.
- It captures the time-varying conditional correlations between yields at different maturities — most notably the well-documented breakdown of correlation in high-volatility periods.
- It reproduces the negative Campbell-Shiller slopes that single-regime CIR cannot, providing a no-arbitrage rationalization of empirical EH violations.
- The estimated regimes are economically interpretable: one regime is high-volatility, low-mean-reversion, and concentrated in NBER recessions; the other is low-volatility and dominant in expansions. Regime classification correlates with US business cycles.
- The log-linear bond price approximation is shown to introduce only a small pricing error within the maturities studied.
Limitations
- Only two regimes — richer dynamics (more regimes, persistent vs. transitory regimes, asymmetric chains) are not explored.
- Constant transition matrix: regime transitions do not depend on macroeconomic variables, so the model cannot endogenize the regime onset.
- Log-linear approximation: bond pricing relies on a linearization whose error grows with maturity and with the spread of regime parameters; long-maturity bonds are not formally validated.
- Single-factor short rate: extensions to multi-factor square-root models with regimes are left open.
- EMM relies on the SNP auxiliary model being a sufficiently good density approximation; misspecification in the auxiliary model contaminates the structural inference.
- Markov chain is latent: the estimated regime classification is conditional on the structural model and is not directly observable.
Open questions
- Can the regime mechanism be linked structurally to macro variables (output gap, inflation, monetary policy stance) instead of being treated as a latent chain?
- How does the regime-switching CIR mechanism interact with multi-factor affine term structure models (e.g., Dai-Singleton-Yang)?
- Do regime shifts also resolve the bond risk premia “puzzles” beyond the Campbell-Shiller violations (e.g., excess return predictability via the Cochrane-Piazzesi factor)?
- Are there better estimation strategies than EMM (e.g., Hamilton filter likelihood, particle filter, MCMC) that exploit the no-arbitrage structure more directly?
My take
This paper is a foundational template for the entire “regime-switching no-arbitrage” branch of term-structure modeling and is directly load-bearing for the CRE asset-pricing project. The mechanism — let drift, volatility, and the price of risk all switch with a latent Markov chain, then preserve no-arbitrage within each regime — is exactly the architectural choice we adopt for the 4-compound-regime monetary-policy × wage-rigidity NK + asset-pricing model. Three lessons transfer directly:
- Joint identification matters: EMM (or any moment-based estimator) requires that the auxiliary model be expressive enough to tease apart regimes from conditional heteroskedasticity. In our project, RBPF plays the analogous role of “expressive auxiliary likelihood” for the latent regime path.
- Log-linear bond approximations have a cost: the paper documents this cost as small for short maturities but growing with maturity. Our
research/pricing_approximation/arc independently confirmed a 13–17% structural bias for the M0 log-linear approximation, consistent with the Bansal-Zhou caveat. - Regime interpretability is a real asset: the paper’s filtered regimes line up with NBER cycles, which is what gives the model its economic credibility. We should look for the same kind of post-hoc interpretation in the CRE incumbent’s regime path.
What the paper does not do, and what our project must extend: it does not connect the regime to a structural macro model (NK, Taylor rule, monetary policy chain), and it does not handle multi-factor or multi-asset extensions.