A GDP-Weighted Multi-Signal System for Real-Time Recession Detection

We build and pre-register a three-signal recession detection system from public administrative data: weekly state-level unemployment insurance claims, GDP-weighted so that the cross-section answers how much of the economy is located in deteriorating states, with a sequential confirmation rule coupling the national and propagation signals and an industry layer that diagnoses where each recession began. The paper is simultaneously a capability study and an honest re-test: the system descends from a pinned manuscript claiming its default detected all eight post-1967 recessions ahead of the Sahm Rule with fewer false positives, and we re-adjudicate that claim under a pre-registered rule on a freshly frozen vintage. The verdict is REFUTE: the default detects seven of eight — the 2001 recession is undetected by a vintage-fragile confirmation-window knife edge — and the honest false-positive comparison runs 4 system episodes to 2 Sahm episodes under identical accounting. The miss is settings-bound, not architectural: 36 of 54 pre-registered pairings detect all eight, every miss anywhere in the space is 2001, and the zero-false-positive full-detection share is 13% — structural, though far below the source's claimed 76%. What the measured system delivers: a quantified earliness–cleanliness frontier (zero false positives and positive mean lead never coexist); a convergence-speed diagnostic that reads recession propagation type from the signals' timing at no added cost; epicenter identification where a broad industry origin exists — for 2007, Real estate (footprint 0.870737 across 46 declining states), corroborated across independent layers, with per-cell footprints, margins, and runner-up rows disclosed so readers can weigh the statistic's known artifact modes — and a plain statement where none does (2001); and a concrete, low-cost data-infrastructure agenda that would relax the system's binding limits. Deployment is reframed as a three-tier posture ending in the Sahm Rule as independent confirmation, and the paper closes by registering that posture as a dated, falsifiable forward claim from a live all-quiet board (Stage 1 QUIET, Stage 2 QUIET at the frozen vintage), with every load-bearing number bound to committed code and hashed data by a public machine-verified ledger. Verified rebuild under the Research-to-Publication Standard v1.9.12: every load-bearing number is registered in a machine-checked ledger (claims.lock) and regenerated on demand by verify.py from hash-pinned public administrative inputs, then independently reproduced from a clean checkout with every experiment re-run in tolerance; a capped single-round adversarial review under a fix-or-rebut protocol and a public corrections log are committed in the repository. A dated, reader-runnable forward prediction is registered publicly as a prospective out-of-sample falsifier — that ahead of the next NBER-dated U.S. recession's announcement, the registered three-tier ladder escalates in order: Stage 1 WATCH, the 10%/4-week national configuration confirming the 25%/8-week propagation configuration under the paper's coupling rule; Stage 2 ELEVATED, the 18%/3-week national configuration confirming the same propagation configuration; Stage 3 CONFIRMED, the Sahm Rule at its public 0.50 trigger — or fails in the concentrated-onset, archetype-conditional way the paper measures, with a Stage-1 alert that Stage 2 never joins counting as a self-cancelling disconfirmation. The prediction is resolved on the same public series the paper uses — state-level unemployment insurance claims and the published Sahm indicator — by the committed pipeline, from the registered all-quiet starting board frozen at the paper's data vintage, and carried forward until the next NBER-dated recession is announced. The paper is licensed CC BY-NC-ND 4.0; the accompanying plain-English companion is CC BY-NC 4.0, and the analysis and verification code is MIT-licensed.