Out-of-sample CRE cap rate forecasting at 1Y+ horizons under regime switching

Motivation

The Leather-Sagi MS-RE model has been validated in-sample on simulated DGP data (local recovery PASS, 5/5, CRE RMSE 2-3%), but out-of-sample forecasting performance — the ultimate test of a structural model’s economic content — has not been evaluated. Meanwhile, the REHaystack (2025) finding that simple reduced-form cap rate forecasters (mortgage debt flows + lagged cap rates + unemployment) achieve R-squared ~0.96 at 1Q but collapse to 0.17-0.31 at 1Y creates a clear opening: the regime-switching structural model encodes forward-looking information about monetary policy paths that reduced-form models cannot capture, and this information should matter most at longer horizons.

Two weakly-supported claims are directly at stake:

• anticipated-monetary-policy-regimes-drive-cre-cap-rates (conf 0.65): if anticipated regime changes drive cap rates, the model should forecast cap rate movements around regime transitions better than reduced-form alternatives.
• cre-risk-premia-drive-cap-rate-cycles (conf 0.70): if risk premia dominate cap rate cycles, the structural model’s regime-switching risk premia should predict cap rate movements better than cash-flow-based models.

Hypothesis

The MS-RE structural model produces superior out-of-sample CRE cap rate forecasts at 4Q and 8Q horizons relative to (a) random walk, (b) reduced-form macro predictors, and (c) Hamilton-only (no cap rate) estimation — with the advantage concentrated around periods of monetary policy regime transitions.

Approach sketch

1. Expanding-window estimation: estimate the model on 1992Q1 through t, forecast cap rates at t+1Q, t+4Q, t+8Q. Re-estimate quarterly or annually.
2. Forecast competitors:
   • Random walk (cap rate stays constant)
   • Reduced-form: debt flows + lagged cap rates + unemployment (REHaystack)
   • VAR(1) on macro variables + cap rates
   • Hamilton-only model (yields but no cap rates in observation menu)
   • Full RBPF model (yields + cap rates)
3. Evaluation metrics: RMSE, MAE, directional accuracy, Diebold-Mariano test of equal predictive ability.
4. Regime-conditional analysis: split forecast evaluation by the model’s smoothed regime probabilities — does the structural model have its largest advantage in periods where the regime is transitioning?
5. Staging: start with simulated-DGP forecast validation (pseudo-out-of-sample), then extend to real NCREIF/Green Street data once the global optimization pipeline is available.

Expected outcome

The structural model matches or beats reduced-form competitors at 1Q and dominates at 4Q+ horizons, with the advantage largest in regime-transition periods. If confirmed, this elevates both target claims and provides the first published out-of-sample forecasting evaluation for a no-arbitrage MS-RE CRE pricing model.

Risks

• Real-data estimation requires the global optimization pipeline (idea global-optimization-pipeline-msre-cre) to be operational. Can be staged with simulated-DGP forecasting first.
• The expanding-window approach is compute-intensive: each re-estimation requires a full RBPF optimization (~350 ms/eval x many optimizer evaluations).
• If the model’s in-sample fit is driven by parameter flexibility rather than economic structure, out-of-sample performance may not improve over simple benchmarks — this would be a negative but informative result.
• Data availability: updated NCREIF/Green Street data through 2025 may be needed.

Pilot results

(empty)

Lessons learned

(empty)