Each subdirectory below is a dated research bundle linked from an indicator piece's validation_artifact field. These are working analysis outputs — CSV results tables, model scripts, structured findings — produced by Rate Authority's empirical validation pipeline. Their contents and conclusions are summarized in the published indicator pieces; this directory exposes the underlying data for replication.
V1 exploratory correlation analysis. Pairwise residualized Spearman + Pearson at lags 0–24 months across BLS CPI/PPI subseries, NOAA weather counts, FRED macro. Random Forest permutation importance. 50-row triple scan. Headline: CPI Motor Vehicle Parts at 12-month lag (ρ=+0.49, n=279). Cited as exploratory; not pre-registered.
→ findings.md · inventory.md · pairwise_results.csv · rf_importance.csv · triples_results.csv
V2-C disciplined re-validation of V1 findings. Pre-COVID stability check (2001–2019 subsample), Granger causality tests, triple/quad extensions. Headline: 7 of 10 V1 findings stable pre-COVID; CPI Motor Vehicle Parts at 12m lag STRENGTHENS pre-COVID (ρ goes 0.47 → 0.54). Zero Granger graduations at p<0.01 due to 12-24m test-lag ceiling. Best quad adj-R²=0.593.
→ validation_log.md · pre_covid_stability.csv · triples_top50.csv · quads_top12.csv · granger_results.csv
BOTH GATES PASS. Gate 6 (full confounder residualization vs Unemployment lag-24, CPI Shelter, CPI Medical, Fed Funds Rate): residualized ρ=0.485 with RMSE skill 0.165 — primary kill mode (UNRATE@24m absorbing the signal) did NOT fire, UNRATE@24m independently fails at ρ=−0.066. Gate 7b (alternative-outcome replication): Transportation Commodities NEC ρ=0.548, CPI Less Food & Energy ρ=0.363, CPI Used Cars correctly fails to replicate (ρ=0.00 — confirms signal specificity to insurance/inflation, not vehicle-price cycles). 8-gate status: 4 pass / 0 fail / 4 pending.
→ validation_extensions_log.md · gate6_residualized_correlations.csv · gate7b_alternative_outcomes.csv · run_gate6_7b.py
Gate 1 FAIL (calibration), Gates 2 + 5 PASS. Brier walkforward (train 2002–2018, test 2019–2024, n=72): test-window Brier=0.1281 vs 0.10 pass threshold — fails on calibration precision, but well below 0.20 kill threshold. Continuous OOS Spearman ρ=0.8551 — direction OOS is extremely strong. Gate 2 BSS=0.4875 (48.7% reduction in Brier loss vs naive 50/50). Gate 5 conviction subset (|p−0.5|>0.20, n=30) Brier=0.0128 with 100% direction accuracy. 8-gate status: 5 pass / 1 fail / 2 pending. The fail is a calibration recalibration job, not a signal-absence kill.
→ validation_brier_log.md · gate1_brier_walkforward.csv · gate2_bss_results.csv · gate5_conviction_subset.csv · run_gates_1_2_5.py
Construction-rebuild cost leading indicator model. 20-year BLS PPI panel (lumber, copper, gypsum, construction labor) vs PPIIDC (FRED proxy for Marshall-Swift Reconstruction Cost). Headline: lumber leads PPIIDC by ~1 year (r=0.72); construction labor wages 2-year peak-lead is the novel union-contract-cycle finding. n=51 single-year cross-section caveated.
→ construction_findings.md · inventory.md · construction_cost_data.csv · construction_predictors_ranked.csv
Wildfire-risk leading-indicator model. NOAA ONI (915 monthly records), NIFC national acres, III/CDI/Aon insured wildfire losses. Track A CA-only HistGBM CV R² negative (n=31 fail). Track B multi-state binary CV AUC=0.658 ± 0.092 at n=155 — modest. La Niña→drought→fire is rediscovery; novel framing is compound multi-lag ENSO to insured-loss prediction. 2020 El Niño catastrophe disconfirms ENSO-alone.
→ wildfire_findings.md · inventory.md · wildfire_model_data.csv · wildfire_predictors_ranked.csv
DOI/SERFF scraper diagnostic + curated-ingest run. Rate Authority ledger grew 26 → 216 carrier-level DOI filings across 16 states. 164 records via new direct_ingest.py from carrier press releases + SERFF public bulletins. Live scrapers: CA working (52 records); FL/TX/NY broken (portal redesigns); SERFF Cloudflare-blocked.
All research artifacts are CC BY 4.0. Cite as: Rate Authority. [Artifact name]. https://rateauthority.org/research/[path]/. Pre-registration scopes for hypotheses are maintained internally; available on request at [email protected].