Statistical Fingerprints of Forced and Unforced Variability Reveal Inconsistencies Between Marine Proxies and Climate Models on Multi‐Decadal to Millennial Timescales

Abstract Comparing general circulation models (GCMs) and Holocene paleoclimate proxies reveals large mismatches between simulated and reconstructed low‐frequency temperature variability on multidecadal to multicentennial timescales. These mismatches—stemming from proxy‐related artifacts or model deficiencies in simulating forced or unforced variability—impact future warming predictions and past climate interpretations. Theory and models suggest forced and unforced variability can be statistically fingerprinted based on robust differences in their spatial pattern and magnitude on these long timescales. This motivates us to characterize frequency‐dependent behavior across 52 globally‐distributed Holocene sediment cores with Mg/Ca‐ and ‐based temperature reconstructions and a suite of forced and unforced GCM simulations using two metrics: spectral slope and cross‐coherence. Compared with unforced simulations, forced simulations exhibit steeper and higher cross‐coherence on centennial‐and‐longer timescales, especially in regions with applied forcing. Proxy‐system models suggest Mg/Ca and ‐based temperature reconstructions can preserve some statistical differences between forced and unforced and coherence. We find inconsistencies between models and proxies, with proxies having higher , indicating more variability (forced or unforced), but lower coherence between sites, indicating more unforced variability. Our analyses highlight: (a) coherence among nearby sites as an important measure of the regional and large‐scale information retained in those records, (b) the need for a larger set of long climate simulations to probe the range of model behavior on these timescales, and (c) the need to integrate proxy‐system models with GCM output to evaluate performance of both proxy networks and climate models.