CPI Used Cars Does NOT Lead Insurance Rates — A Publishable Null (2026)

Last updated May 2026 · Rate Authority.

CPI Used Cars Does NOT Lead Insurance Rates — A Publishable Null (2026)

Empirical contradiction or kill-mode fired: CPI Used Cars and Trucks (CUSR0000SETA02) eliminated as a leading indicator of consumer insurance rate proxies

(Source: Rate Authority, May 2026.)

Every commentator covering US auto insurance rates eventually reaches for the Manheim Used Vehicle Value Index. The logic is immediate and intuitive: used-car prices set the salvage-value floor for comprehensive coverage. A totaled vehicle is worth what the salvage market will pay. When used-car prices rise, insurers’ total-loss exposure shrinks — and when they fall, it rises. The Manheim index is real-time, widely available, and routinely cited in trade press as a leading indicator of where auto insurance CPI will go next.

The problem with intuitive leading indicators is that they are rarely tested rigorously. The BLS CPI series for Used Cars and Trucks (CUSR0000SETA02) traces the same underlying asset that Manheim covers. Rate Authority ran the CUSR0000SETA02 series through the same disciplined discovery harness that produced the validated findings published elsewhere on this site. The hypothesis is a clean statement: CPI Used Cars and Trucks, lagged zero to twenty-four months, shows a meaningful positive correlation with consumer insurance CPI proxies.

It does not hold. This piece publishes the kill.

NerdWallet, Bankrate, and MoneyGeek do not publish their kills. When a potential leading indicator fails their internal testing — if they test at all — the failure disappears. The surviving findings are the ones readers cite. Rate Authority operates under the opposite discipline: the kills are published because they are the proof that the surviving findings are real. The Tier B per-break differential layer in the surf-ecosystem cat-bond methodology was killed with 3-of-4 sub-gates passing, and the kill-log was published alongside the validated Tier A result. That is the standard here.

The numbers

The V1 Random Forest scan tested 120 features across all BLS FRED macro series. CPI Used Cars and Trucks (CUSR0000SETA02) at a 24-month lag produced a permutation importance z-score of 9.27 — high enough to advance past the first-stage importance screen into the second-stage correlation validation. This is the canonical false-positive failure mode in multi-feature machine learning: a permutation score can reflect correlated noise in a 120-feature search space rather than a real signal. The second stage is where the kill fires.

Source: /research/v2c_validation_2026-05-22/pre_covid_stability.csv and /research/gate_6_7b_validation_2026-05-22/gate7b_alternative_outcomes.csv.

The ”% change” column in the V2-C output for this row reads −301.8% — a mathematical artifact of the near-zero denominator, not evidence of instability. The honest reading is simpler: the correlation is near zero on both windows, and there is no delta to interpret. A change of −0.074 in absolute ρ terms, moving from −0.025 to −0.099, is noise at these magnitudes.

Compare this to the validated finding on the same run. CPI Motor Vehicle Parts at 12-month lag (CUSR0000SETC) produces Spearman ρ = +0.470 on the full window (n=291) and ρ = +0.540 on the pre-COVID subsample (n=216, Δ = +0.070, 14.8% change, status: stable). The contrast is not ambiguous. One series predicts insurance rates. The other does not.

Why the consensus mental model is wrong

The intuitive argument — used-car prices set the salvage-value floor for comprehensive coverage, therefore used-car prices lead insurance rates — makes a conceptual error in the distinction between salvage values and claim severity.

Comprehensive insurance premiums are driven by the expected cost of covered claims. A total-loss claim on a comprehensive policy does recover the vehicle’s salvage value. But comprehensive claims make up a small fraction of the total claim cost stack. The dominant cost driver in auto insurance is repair-cost severity: the labor hours and parts expense for partial-damage repairs. That cost is captured in CPI Motor Vehicle Parts (CUSR0000SETC, the series that survives the harness), not in used-car transaction prices.

There is a second error in the salvage-value argument. Even for total-loss claims, the insurer’s economic exposure is the difference between the vehicle’s actual cash value and the salvage recovery — not the salvage level in isolation. When used-car prices rise, actual cash values and salvage recoveries rise together. The insurer’s net exposure on a total-loss claim does not necessarily shrink when used-car prices increase; it depends on the relative movement of ACV against salvage. The Manheim index tracks salvage-market prices without adjusting for the simultaneous movement in vehicle ACV. A commentator citing Manheim as a leading indicator of insurance loss trends is not making a net-exposure calculation.

The third error is timing. Comprehensive premiums are recalibrated in the state rate-filing cycle, which runs on its own clock — typically 12 to 24 months behind the underlying cost experience. The CPI Used Cars series is repriced monthly by BLS survey. The two series operate on incommensurable timescales. The CPI MV Parts → insurance-rate lead is approximately 12 months because it runs through repair costs → actuarial loss trends → state filing → premium adjustment → BLS CPI print. The salvage-value channel, even if real, would run through total-loss frequency → severity → actuarial loss trends → state filing → premium adjustment — a longer chain with more noise at every step.

The empirical result is consistent with this structural reading. CPI MV Parts at lag-12 shows Spearman ρ = +0.470 full-window, ρ = +0.540 pre-COVID, RMSE skill = +0.165 after full-confounder residualization. CPI Used Cars shows ρ = −0.025 full-window, ρ = −0.099 pre-COVID, with a Granger causality test returning p = 0.879 at the optimal AIC lag. There is no signal.

Implication for forecasters: stop citing Manheim as an insurance leading indicator for premium rate cycles. The relevant BLS series are CUSR0000SETC (CPI Motor Vehicle Parts, validated at 12-24 month lags) and PPI WPU141 (PPI Motor Vehicle Parts, robustness series). The salvage-floor narrative is intuitive, widely cited, and empirically unsupported at the level of CPI print correlation.

What we learn from the kill

Three findings in the V1-to-V2-C validation cycle needed the pre-COVID stability second stage: CPI Used Cars, Unemployment Rate at 24-month lag, and CPI Shelter/OER at 0-month lag. They resolved differently.

CPI Used Cars died cleanly. Full-window ρ = −0.025, pre-COVID ρ = −0.099. Near zero across both windows. No signal in the normal rate cycle, no signal when the 2020–2022 COVID supply-chain episode is excluded. Kill fires.

Unemployment Rate at 24 months weakened to a caveat. Full-window ρ = +0.320 → pre-COVID ρ = +0.496, Δ = +0.177. The correlation is stronger in the pre-COVID window — the opposite direction from a COVID-inflated false positive. The finding survives as directional but with a mechanism-revision caveat: the unemployment lag-24 correlation may reflect recession-cycle insurance purchasing behavior rather than a structural rate-pass-through mechanism. It is not killed; it is re-scoped.

CPI Shelter/OER showed the opposite pattern. Full-window ρ = −0.293 → pre-COVID ρ = −0.640, Δ = −0.347. The pre-COVID signal is nearly twice as strong as the full-window estimate, which suggests the COVID episode partially obscured the relationship rather than creating it. The mechanism interpretation requires revision — the sign is negative, and the structural story is different from what V1 suggested — but the signal is real.

The value of running all three through the same harness is precisely this: one clean kill, one caveat, one mechanism revision. Without the pre-COVID stability gate, all three would sit in the “directional” bucket with the same apparent status. The kill-log discipline separates them.

A high Random Forest permutation importance score is not evidence of a real signal. CPI Used Cars at lag=24m produced importance z-score = 9.27 — higher than CPI Electricity (z=7.34), Rent (z=8.49), and Food at Home at lag=12m (z=9.70). The permutation importance ranks on predictive contribution in the full multi-feature model, which includes cross-feature interactions and collinear structure. In a 120-feature scan, a near-zero univariate predictor can rank high through its correlation with real predictors. The second-stage residualized Spearman + pre-COVID stability is the correct filter. The first-stage importance screen is the candidate nomination mechanism, not the verdict.

This is the core methodological lesson the kill teaches. The discipline of running permutation importance to nominate, then residualized Spearman to validate, then pre-COVID subsampling to confirm structural stability, then alternative-outcome testing to confirm signal specificity, produces clean kills that the permutation-only workflow would have promoted to publication.

What this kill does not mean

It does not mean Manheim is useless. The Manheim Used Vehicle Value Index is a high-quality, high-frequency series with genuine predictive validity in its own domain. It predicts used-car-market liquidity, dealer floorplan economics, repossession volume, auction throughput, and lender-level collateral valuations. It correctly called the 2021–2022 used-vehicle price spike before BLS CPI reflected it. None of that is in dispute. What this kill says is narrower: Manheim-level used-car price movements, as measured by the BLS CPI Used Cars and Trucks series, do not translate into meaningful leading correlations with consumer insurance CPI proxies at the lags we tested (0 to 24 months, monthly resolution, 2001–2026).

It does not mean salvage value is irrelevant to insurance pricing. Total-loss salvage economics are a real input to comprehensive premium pricing. They are simply not the leading input on the timescale captured by CPI cross-correlation. The rate-filing cycle, the actuarial reserving cycle, and the BLS CPI measurement cycle all insert lags and smoothing that attenuate the salvage-market signal to near zero by the time it appears in CPI print.

It does not rule out a future finding. If BLS reconstructs a dedicated CPI Motor Vehicle Insurance series — CUSR0000SETC01 was the canonical series, now retired from FRED — with a full historical backfill, the analysis would benefit from a direct auto-insurance outcome variable rather than the CPI Tenants/Household Insurance substitute (CUSR0000SEHG) used here. Used-car prices may show a different lag structure against pure auto-insurance CPI than against the tenants insurance proxy. That is a testable hypothesis for a future data cycle, not a claim the current data can adjudicate.

It does not mean the alternate-outcome result is a kill-for-the-predictor. Gate 7b tested whether CPI MV Parts (CUSR0000SETC at lag=12m) correctly fails to predict CPI Used Cars as an outcome, specifically as a check on signal specificity. CPI MV Parts → CPI Used Cars yields residualized ρ = −0.0001 (p = 0.999) — a precise near-zero result that confirms the MV Parts signal is specific to insurance and inflation contexts, not to vehicle-price cycles generally. This is the correct result: a predictor that succeeds on insurance outcomes should not also predict used-car prices, or the insurance signal is just a vehicle-market proxy in disguise. Gate 7b passed, which is evidence for the MV Parts finding, not evidence against anything.

The distinction that matters

Two BLS series, both in the “motor vehicle” family, tested against the same insurance CPI outcome, same methodology, same window:

The kill is not a modeling failure. It is the model working correctly. The method that kills CPI Used Cars is the same method that validates CPI Motor Vehicle Parts. That is the proof the methodology has power to discriminate.

Citation and methodology

Methodology: BLS CPI series via FRED (CUSR0000SETA02, CUSR0000SEHG, CUSR0000SETC, CUSR0000SETD). Residualization on month-of-year fixed effects per Rate Authority standing protocol. Eight-gate discovery harness applied: V1 Random Forest permutation importance (120-feature scan), Spearman correlation on full 2001–2026 window (n=279), V2-C pre-COVID stability subsample (2001–2019, n=204), Granger causality bivariate test (optimal AIC lag, F-test p<0.01 threshold), gate 6 residualized RMSE skill, gate 7b alternative-outcome test (CPI MV Parts → CPI Used Cars as outcome, confirming used-car series correctly fails to replicate the insurance-specific signal). Full V2-C run: /research/v2c_validation_2026-05-22/pre_covid_stability.csv. Gate 7b run: /research/gate_6_7b_validation_2026-05-22/gate7b_alternative_outcomes.csv.

Internal cross-references:

• What works instead: CPI Motor Vehicle Parts as a validated 12-month leading indicator
• Full signal scan methodology: Cross-signal correlation analysis, 2026
• Gate definitions and conviction tiers: Rate Authority conviction-tier methodology
• Underlying data run: V2-C validation, 2026-05-22

Cited as: Rate Authority. CPI Used Cars Does NOT Lead Insurance Rates — A Publishable Null (2026). https://rateauthority.org/indicators/cpi-used-cars-killed-as-leading-indicator-2026/.

Maintained by Rate Authority Editorial. Operated by PolicyChat.

Methodology: Rate Authority’s confidence-tier framework — see /methodology/rate-authority/. This piece is tier kill_log. Rate Authority’s editorial decisions and methodology are independent of any commercial relationship.