Regimes, Not Forecasts: Reassessing Dynamic Nelson-Siegel Term Structure Forecasting, and a Proposed Descriptive Alternative

Diebold and Li (2006) showed that a simple autoregressive model of the Nelson-Siegel yield curve factors produces genuinely useful 12-month-ahead forecasts. Replicated on an extended U.S. Treasury sample running through 2026 — roughly 26 years beyond their original 1985-2000 window — that result does not hold: the model loses to a naive random-walk forecast on every tenor tested, under both expanding and rolling estimation windows, with degradation frequently exceeding 10%. The mechanism is identifiable and consistent with independent structural work: the zero-lower-bound period introduced a materially different factor dynamic that the forecasting specification could not accommodate. Rather than treat this as a dead end, this paper develops and formally tests an alternative: a small number of recurring, economically interpretable curve-shape states — fewer in number than they first appear, once rate level and curve shape are properly separated. Seven falsifiable hypotheses are tested: bootstrap cluster stability, permutation-tested temporal persistence, chi-square correspondence with Fed policy phase and NBER recession dating (with effect sizes reported), a formally BIC-gated comparison showing a unified hidden Markov model outperforms static clustering (robust to leave-one-episode-out testing), and surrogate-tested spectral hypotheses returning informative negative results. Three analytical errors identified and corrected during testing are reported explicitly. The resulting framework is deliberately descriptive rather than predictive, and is extended into an operational monthly monitoring tool built on causal (filtering-based) inference rather than retrospective smoothing. The paper's central claim is narrower than a forecasting model would be: the curve's useful information may lie in its current state rather than its future trajectory.