Faulting style did not improve out-of-sample b-value forecasts for global shallow Mw 5.7+ GCMT earthquakes in 2004-2026, but plate-boundary setting did

Schorlemmer et al. (2005) found that the Gutenberg-Richter b-value is highest for normal faults, lowest for thrust faults and between for strike-slip faults. We tested this relation out of sample in the Global CMT catalog. We fitted b as a linear function of the folded rake for shallow (centroid depth 50 km or less) Mw 5.7+ earthquakes in 1977-2003 (4,726 events) and tested the frozen model on 2004 to mid-2026 (4,838 events). The plan was written before the b-rake statistics. The training slope was -0.138 per 90 degrees of rake (95% tile-bootstrap CI -0.203 to -0.072). The test slope was -0.073 (-0.129 to -0.022), half as large. The frozen rake model gave no test log-likelihood gain over a constant b (+0.0002 nats per event, CI -0.0037 to +0.0029), so the pre-set transfer rule failed. With one b intercept per Bird (2003) plate-boundary setting, the test slope was -0.005 (-0.048 to +0.034); in training it was -0.098. Spreading ridges and oceanic transforms have high b and few thrust events; subduction zones have b near 1.0 and most thrust events. Inside subduction zones there was no rake slope in either period. In an exploratory comparison, a frozen setting model gave +0.0078 nats per event (CI +0.0032 to +0.0125) in the test period, and rake added nothing. The within-setting slope was negative in all three pre-2004 windows and near zero in all three later windows. This step coincides with GCMT procedure changes in 2004, which we did not test as the cause. Limitations: one catalog, Mw 5.7+ only, a coarse nearest-boundary setting label, strong rake-setting correlation, and a low-power test of the corner-magnitude effect. Code, plan, aggregate tables and figures are included; raw catalogs are retrieved by script.

b-value · Gutenberg-Richter · faulting style · rake · Global CMT · plate boundaries · PB2002 · out-of-sample test · differential stress · earthquake magnitude distribution · replication · catalog homogeneity