Predicting Dividend Cuts in U.S. REITs under Interest-Rate Shocks: A Regime-Gated Monotone Additive Model with Conformal Risk Control

Abstract

Interest-rate shocks reach the distributions of real estate investment trusts (REITs) through asset revaluation, debt repricing and refinancing frictions, but the resulting cut risk is difficult to model. It is conditional: in our data a rate shock raises the four-quarter cut probability of the most rate-exposed quartile of REITs from 16.6% to 19.8% while barely moving the least exposed quartile, so a pooled regime dummy cannot recover the channel. We propose RG-MAM, a regime-gated monotone additive model in which every feature has a closed-form shape function built from non-negative combinations of monotone ramp bases, plus a gated increment that is active only in rate-shock quarters and is itself sign-restricted. The explanation is therefore the model, not a post-hoc approximation, and the response to a parallel rate shock is provably non-decreasing. An ordinal head with ordered cut-points extends the guarantee to first-order stochastic dominance of the severity distribution. We add RA-CRC, a regime-adaptive conformal procedure that restores out-of-time recall control when exchangeability fails. On 24,693 firm-quarters built from observed U.S. macro-financial series (1990-2025) and an accounting-consistent REIT panel, RG-MAM attains out-of-time AUC 0.722 with expected calibration error 0.033, three times better calibrated than a class-balanced logistic benchmark of statistically indistinguishable ranking. A +300 bp counterfactual raises the mean cut probability from 10.2% to 17.4% monotonically for all 240 REITs and zero ladder steps, whereas an otherwise identical unconstrained model, only 0.015 AUC better, violates monotonicity for every firm.