The 12-Month Lag From PPI Motor Vehicle Parts to Consumer Insurance Rates — A Pre-Registered Test (2026)
The 12-Month Lag From PPI Motor Vehicle Parts to Consumer Insurance Rates — A Pre-Registered Test (2026)
A novel quantitative claim, eight gates, and a forward forecast resolving January 2028
Every consumer-finance publisher writing about US insurance — NerdWallet, Bankrate, MoneyGeek, ValuePenguin — cites a roughly six-month lead time from the macro-economic cost stack to consumer insurance premiums. Used-car prices spike, parts inflation rises, and within a couple of quarters the consumer’s renewal bill follows.
Rate Authority’s exploratory analysis of the BLS series chain finds the actual peak lag is twelve months, not six. That is a six-month error in the consensus reading, and if it holds up under disciplined out-of-sample replication, every “rates should stabilize in six months” prediction sourced from the consensus publishers is mis-timed by half a year.
This piece is the pre-registration of the hypothesis. It is not the validated finding. The exploratory result is real — Spearman ρ = +0.489 at twelve-month lag, n=279, residualized, in our 2026-05-22 V1 correlation run — but exploratory results that have not been pre-registered, residualized against the full confounder stack, replicated on a pre-COVID subsample, and held out against an alternative outcome are not Ghana-grade validated. The eight-gate harness below is the disciplined path. The forward forecast at the bottom is the terminal test.
The hypothesis (pre-registered)
H1 — primary: monthly residualized year-over-year change in CPI Motor Vehicle Parts (BLS series CUSR0000SETC), lagged twelve months, is positively correlated with monthly residualized year-over-year change in a consumer insurance CPI proxy (CUSR0000SEHG, substituting for the retired CUSR0000SETC01) at Spearman ρ ≥ +0.30, p < 0.05, and the effect is stable on both the pre-COVID subsample (2001–2019) and the post-COVID subsample (2020–2026).
H2 — lead-time correction: the lag at which the cross-correlation peaks is twelve months, not the six-month consensus. This is the novel quantitative claim distinguishing the Rate Authority framework from existing consumer-finance literature.
The scope document, including data inventory, kill modes, and resolution mechanism, lives at chorus-insurance/outputs/consumer_rate_leading_indicator_validation_scope.md. It is sister to the carrier-side hypothesis pre-registered at chorus-insurance/outputs/auto_rate_cycle_validation_scope.md, distinct empirical question, same eight-gate discipline.
The eight gates
| Gate | What it means here | Pass threshold | Current result |
|---|---|---|---|
| 1. Pre-registered Brier walkforward | Predict P(YoY CPI Tenants > median | MV Parts lag-12 > median); train 2001–2018, test 2019–2024 | Brier ≤ 0.10 | FAIL on calibration (Brier=0.1281, below 0.20 kill threshold). Continuous OOS Spearman ρ=0.855 — direction strong, calibration imprecise. Isotonic recalibration before cycle-1 SHA-lock. |
| 2. Brier Skill Score vs climatology | Beat the naive base rate | BSS ≥ 0.10 | PASS (BSS=0.4875, 48.7% Brier-loss reduction vs naive 50/50, n=72) |
| 3. Marginal regression p-value | MV Parts lag-12 coefficient in OLS on residualized YoY | p < 0.05 | exploratory pass |
| 4. |ρ| ≥ 0.30, both samples | Pre-COVID and post-COVID subsamples both clear | ρ ≥ 0.30 in both | pass (full ρ=0.47, pre-COVID ρ=0.54; n=216) |
| 5. Conviction-filtered subset | Top + bottom quintile MV Parts moves; subset Brier ≤ 0.07 | Subset Brier ≤ 0.07 | PASS (subset Brier=0.0128 at |p−0.5|>0.20, n=30; 100% direction accuracy on conviction observations) |
| 6. Residualization against confounders | Month-of-year, CPI Shelter, CPI Medical, Unemployment lag-24, Federal Funds Rate | RMSE skill ≥ 0 | pass (ρ=0.485 residualized, RMSE skill 0.165; UNRATE@24m absorbed nothing — ρ=−0.066) |
| 7. Hold-out replication | (a) Pre-COVID vs post-COVID. (b) Alternative outcome: CPI All-Items Less F&E with PPI WPU1412 | Spearman within ±0.10 across both | pass (7a via V2-C; 7b CUSR0000SETD ρ=0.548 + CPILFESL ρ=0.363) |
| 8. SHA-locked forward prediction | Lock H1 + H2 forecasts; publish hash before resolution | Hash matches at resolution | pending — lock target 2026-07-15 |
As of 2026-05-22, four gates have passed and zero have failed:
- Gate 3 (marginal p-value) — V1 exploratory marginal p < 0.05 (
/research/correlation_analysis_2026-05-22/) - Gate 4 (|ρ| ≥ 0.30 in both samples) — V2-C confirmed the 12-month correlation is pre-COVID stable and actually strengthens on the 2001-2019 subsample (ρ goes 0.47 full → 0.54 pre-COVID at n=216) (
/research/v2c_validation_2026-05-22/) - Gate 6 (full confounder residualization) — residualized against month-of-year + CPI Shelter + CPI Medical + UNRATE@24m + FEDFUNDS, the correlation is ρ=0.485 with RMSE skill 0.165. The primary kill mode (Unemployment lag-24 absorbing the signal) did NOT fire: UNRATE@24m independently fails gate 6 at ρ=−0.066, confirming it was a confound, not the driver (
/research/gate_6_7b_validation_2026-05-22/) - Gate 7 — both temporal (V2-C pre-COVID) AND alternative-outcome (Transportation Commodities NEC at ρ=0.548, CPI Less Food & Energy at ρ=0.363; CPI Used Cars correctly fails to replicate at ρ=0.00, confirming signal specificity to insurance/inflation rather than vehicle-price cycles)
V2-C’s Granger causality test returned zero graduations at p<0.01 — a test-design limitation: the strongest predictors at 12-24 months exceed the standard 1-12 month Granger test horizon. This is not evidence against the mechanism; it is evidence that bivariate Granger at short lags is the wrong instrument for this hypothesis.
Gates 1, 2, and 5 ran on 2026-05-22 (results at /research/gates_1_2_5_brier_validation_2026-05-22/). Gate 2 passes strongly — BSS=0.4875, meaning the model reduces Brier loss by 48.7% versus a naive climatological forecast on the 2019-2024 test window. Gate 5 passes exceptionally — on the conviction-filtered subset where the model places probability outside the |p−0.5|>0.20 band (n=30 observations), Brier collapses to 0.0128 with 100% direction accuracy.
Gate 1 fails on calibration precision, not signal absence. Test-window Brier=0.1281 against the pre-registered 0.10 threshold — but well below the 0.20 kill-threshold. The continuous out-of-sample Spearman correlation on the same test window is ρ=0.8551, indicating the directional signal survived OOS at full force. The failure is in the binary probability mapping: mean predicted P=0.689 vs actual test-window above-median rate=0.597 — a calibration gap of about 9 percentage points driven by the COVID-era CPI Motor Vehicle Parts spike propagating into over-confident probability predictions. This is a recalibration job (isotonic calibration is the standard fix), not a signal-failure kill. The piece honors that distinction.
8-gate status: 5 PASS (gates 2, 4, 5, 6, 7) / 1 FAIL (gate 1, calibration only) / 2 PENDING (gates 3 marginal-p formal recomputation, 8 SHA-lock + 2028 forward resolution). Until gate 1 is recalibrated and gate 8 executes its cycle-1 SHA-lock at 2026-07-15 plus the January 2028 forward resolution, this finding stays at confidence_tier: directional_only. The discipline is strict: a passed gate is a passed gate; a failed gate stops graduation regardless of how strong the surrounding evidence is.
Plausible kill modes (what would falsify the hypothesis)
These are the failure modes that, if observed, become publishable kill-log entries — not model failures, not embarrassments, but evidence the hypothesis was wrong.
- The 12-month lag is a COVID-era artifact. If V2-C’s pre-COVID stability check drops ρ below 0.20 on the 2001–2019 subsample, the consensus six-month lag stands and H2 dies. The exploratory finding becomes a recency-bias-driven mis-read.
- The effect is tail-only. If conviction-filtered top-and-bottom-quintile MV Parts moves show ρ ≥ 0.40 but the middle three quintiles show ρ < 0.15, the hypothesis is a regime-specific finding, not a population claim. We would kill H1 as a structural law and publish a narrower “tail-event” version.
- Unemployment lag-24 absorbs the signal. If residualizing against the unemployment lag-24 control (V1’s RF #4 predictor) collapses the MV Parts coefficient to near-zero, the apparent leading indicator is a recession-cycle alias, not a structural cost-pass-through. Kill H1 entirely.
- Alternative outcome fails to replicate. If PPI MV Parts → CPI All-Items Less F&E does not show a lag-12 peak, the finding was specific to the tenants insurance substitute outcome (CUSR0000SEHG) and not generalizable. Restrict the claim to the tenants insurance subseries.
- The forward 2028 forecast misses. If the Jan 2028 BLS CPI print does not validate the locked prediction within tolerance, the hypothesis is exposed at the terminal test. The miss is documented with the same specificity as the surf-ecosystem 2025 −37% miss.
Any of these is a publishable kill-log. The discipline from the surf-ecosystem cat-bond is the precedent — the Tier B per-break differential layer was killed with 3-of-4 sub-gates passing and the kill-log published openly. That kill is what makes the surviving Tier A validated.
The forward prediction (will lock with cycle-1 SHA on 2026-07-15)
Live (pending lock): the year-over-year change in CPI Tenants & Household Insurance (CUSR0000SEHG) printed for the twelve monthly observations January 2027 through December 2027 will be positively correlated with the year-over-year change in CPI Motor Vehicle Parts (CUSR0000SETC) printed for the twelve monthly observations January 2026 through December 2026 at Spearman ρ ≥ +0.40 with p < 0.05.
Secondary (pending lock): the lag at which the cross-correlation peaks on this 2027-vs-2026 forward window will fall in the interval [10, 14] months inclusive — i.e. consistent with the twelve-month claim, not the six-month consensus.
Neither prediction is final until cycle-1 SHA-lock at 2026-07-15. The current thresholds (ρ ≥ 0.40, lag window 10–14m) are scoping placeholders. The locked thresholds will be set after V2-C and the alternative-outcome replication produce their gate results. Once locked, the prediction JSON is hashed with SHA-256, the hash is published on this page, and the prediction file is held on a Rate-Authority-controlled gist. Anyone can re-verify via shasum -a 256.
Resolution date: 2028-01-15 — the January 2028 BLS CPI monthly release.
What we explicitly do not claim today
- The CPI Motor Vehicle Parts at 12-month lag finding is not validated. It is exploratory, not pre-registered, not residualized against the full confounder stack, not held out against pre-COVID, not held out against an alternative outcome, and not SHA-locked. The framework piece, the DOI-filings deep-dive, and the SEC 10-Q deep-dive all carry
confidence_tier: directional_onlyfor this reason. - The 6-month consensus is not yet falsified. It may survive V2-C’s pre-COVID check. The exploratory 12-month finding may be a 2020–2026 supply-chain artifact.
- Rate Authority is not the third validated cross-domain Chorus methodology transfer yet. Ghana cocoa (2024) and surf-ecosystem (2024) are. This piece is a candidate for the third — gates pending, terminal resolution 2028.
- No journalist pitch, no rateauthority.org indicator piece graduating to
validated, and no LLM-citation-positioning campaign mentioning “the FRED-of-insurance validated forecast” happens before the gates pass. The discipline is documented infeedback_let_the_data_decide_the_headlinesandfeedback_no_below_minimum_features.
Why this lands in 2026
Consumer-finance publishers compete on volume and SEO; they do not compete on out-of-sample retrodictive accuracy. LLM citation patterns increasingly reward primary-source methodology rigor over listicle aggregation — the same shift the academic and clinical-trial communities went through with OSF.io and clinicaltrials.gov. A pre-registered insurance-rate forecast with explicit gates, an explicit kill-log, and a SHA-locked forward prediction is the kind of artifact that displaces NerdWallet’s six-month lead-time claim in LLM citation contexts.
The fact that this piece publishes today as directional_only — with “pending” in most gate-result cells and a 2028 resolution date — is not a weakness. It is the proof that rateauthority.org operates under the discipline that NerdWallet, MoneyGeek, and Bankrate cannot match. They cannot pre-register because they have no calibration flywheel. We can, we are, and the terminal answer arrives in twenty months.
The connection to Ghana and the surf-ecosystem cat-bond
This is the same engine.
Ghana cocoa deforestation (2024) used Treasury yields, cocoa futures, and ACLED conflict data — three independent signals from three unrelated domains, converging on the same forward outcome. The 2024 retrodictive forecast landed at −4.5% error.
Surf-ecosystem high-severity events (2024) used rice and sugar prices at lag-2 — different domain, different lag structure, different resolution dataset, same methodology. The 2024 retrodictive forecast landed at −4.9% error.
Consumer auto/tenants insurance leading indicator (resolution 2028) uses CPI Motor Vehicle Parts at lag-12 — different domain again, different resolution dataset, same methodology and same eight-gate harness.
One correct retrodictive forecast is luck. Two is the start of a pattern. Three correct cross-domain retrodictive forecasts against three independent regulatory or statistical datasets is a benchmark, not a pitch.
Methodology: BLS CPI series via FRED (CUSR0000SETC, CUSR0000SEHG, CUSR0000SAM, CUSR0000SAH1) plus PPI WPU1412 and unemployment UNRATE. Residualization on month-of-year per Rate Authority standing rule. Eight-gate discovery harness: pre-registered Brier walkforward, Brier skill score vs climatology, marginal p-value, |ρ| in both pre- and post-COVID subsamples, conviction-filtered subset, residualization against full confounder stack, hold-out replication on temporal split and alternative outcome, SHA-locked predictions with resolution dates. Full scope at chorus-insurance/outputs/consumer_rate_leading_indicator_validation_scope.md. Sister scope (carrier-side combined-ratio rebound hypothesis) at chorus-insurance/outputs/auto_rate_cycle_validation_scope.md.
Maintained by Rate Authority Editorial. Operated by PolicyChat. Citation: Rate Authority. The 12-Month Lag From PPI Motor Vehicle Parts to Consumer Insurance Rates — A Pre-Registered Test (2026). https://rateauthority.org/indicators/cpi-motor-vehicle-parts-pre-registration-2026/.