Cross-Signal Correlation Analysis: What Predicts US Consumer Insurance Rates (2026 Findings)

Cross-Signal Correlation Analysis: What Predicts US Consumer Insurance Rates (2026 Findings)

The consensus is mis-timed by six months

Every major consumer-finance publisher writing about US insurance rates cites the same pipeline: when repair-cost inflation rises, consumer premiums follow within roughly six months. The claim appears on NerdWallet, Bankrate, MoneyGeek, and ValuePenguin in nearly identical language. It is directionally correct. But the consensus has the timing wrong.

Rate Authority’s V1 correlation analysis, launched 2026-05-22 against 25 years of BLS FRED data, found the peak predictive lag between CPI Motor Vehicle Parts inflation and consumer insurance rate proxies is twelve months, not six. The V2-C pre-COVID stability validation, run on the same day against the 2001–2019 subsample alone, strengthened that finding: the correlation does not collapse when you remove the 2020–2022 COVID supply-chain distortion. Pre-COVID ρ at 12 months is 0.540 — higher than the full-window estimate of 0.470, not lower.

The implication is direct. If the actual peak lag is twelve months, every “rates should stabilize by mid-year” prediction based on the six-month consensus is mis-timed by half a year. Rate buyers planning renewals, journalists interpreting monthly CPI prints, and analysts modeling rate cycles are all operating on a stale timeline.

This piece synthesizes the four analyses completed on 2026-05-22 — V1 correlation discovery, V2-C pre-COVID stability and Granger testing, V2-construction cost modeling, and V2-wildfire modeling — and lays out which findings survived disciplined re-validation and which did not.

What this is: a synthesis of exploratory and semi-confirmatory analyses. It is not a validated finding. None of the signals below have cleared the full eight-gate Rate Authority harness. The specific status of each gate is documented in the pre-registration scope.

What this is not: a claim that we have identified causal leading indicators. We have identified correlates that survive pre-COVID stability checks. Granger causality did not graduate a single pair at p < 0.01. That distinction matters and is spelled out in the kill-log section below.

The headline finding

All correlations are Spearman ρ, residualized on month-of-year fixed effects. Full window: 2001–2026 (n = 279–303 monthly observations). Pre-COVID window: 2001–2019 (n = 204–228 monthly observations). Y variable: CPI Tenants/Household Insurance (CUSR0000SEHG) year-over-year change. CPI Motor Vehicle Insurance (CUSR0000SETC01) is retired from FRED and unavailable; CUSR0000SEHG is the substitute.

Source: /research/v2c_validation_2026-05-22/pre_covid_stability.csv. V2-C validation log at /research/v2c_validation_2026-05-22/validation_log.md.

The CPI Motor Vehicle Parts signal at 18m lag (pre-COVID ρ = 0.548) is the single most defensible pairwise finding in the dataset. It strengthened slightly from V1’s full-window estimate of 0.501 when COVID is removed. That is the pattern you want to see from a structural signal: stability or modest improvement under a harder test.

The Shelter/OER reading requires interpretation. The direction is negative — higher shelter inflation is associated with lower insurance rate growth — and it gets substantially stronger pre-COVID (−0.640 vs −0.293 full window). This is counterintuitive for the replacement-cost mechanism but consistent with the hypothesis that Shelter CPI tracks low-catastrophe, low-inflation environments where rate pressure is muted. The signal passes pre-COVID directional stability; the mechanism hypothesis requires further work.

Best linear models (full window, 2001–2026):

Best triple: VehicParts@18m + Food@6m + Unemp@24m → adj R² = 0.499 (with interaction terms; n = 278).

Best quad: Shelter@0m + VehicParts@24m + Electricity@12m + Food@6m → adj R² = 0.593 (main effects; n = 278).

All model R² values are full-window estimates. The COVID supply-chain episode inflates every cost-inflation combination. These should be treated as upper-bound scaffolding, not publishable model fits, until pre-COVID sub-sample validation runs on each finalist.

What the data shows

The PPI-to-CPI auto-parts chain is the backbone

The dominant signal across all four analyses is the Motor Vehicle Parts inflation chain. The mechanism is mechanically grounded: auto body repair is the largest single cost component in personal auto claims. When parts prices rise — whether from supply-chain disruption, tariffs, or materials inflation — claims severity inflates within 2–3 quarters. Actuaries embed the trend in rate filings. State DOI review and approval takes 3–9 months. CPI prints 1–3 months after effective dates. The complete pipeline from PPI shock to CPI rate confirmation spans 4–8 quarters — consistent with a peak lag of 12–18 months, not the 6-month popular consensus.

The V1 RF permutation importance confirms: CPI Motor Vehicle Parts at 12m is the #1 feature (permutation importance = 0.097 ± 0.010), and at 18m is #2 (0.063 ± 0.005). Together they account for approximately 40% of total positive permutation importance across all 120 features. The V2-C pre-COVID stability check confirms neither is a COVID-era artifact: the correlation strengthened slightly under the harder test.

The V2-C validation log notes: “consistent with a state-filing-approval pipeline extension of approximately 6 months beyond the commonly assumed repair-cost-to-rate lag.” That annotation is the most defensible novel contribution of V1. The six-month consensus conflates two distinct pipeline steps: (1) the repair-cost-to-filing lag (2–3 quarters, where the consensus is roughly right) and (2) the filing-to-CPI-print lag (an additional 2–3 quarters that the consensus absorbs but does not separately attribute). The full-chain peak is therefore 12 months, not 6.

Food@Home at 6 months — a partial surprise

CPI Food at Home at 6m lag ranks third by RF permutation importance (0.040 ± 0.004) and has pre-COVID ρ = 0.444 — stable through the COVID exclusion test. Two mechanism hypotheses, neither fully validated:

First, income compression. When food inflation is elevated, household budgets compress. Lapse risk rises as price-sensitive policyholders drop coverage. Adverse selection degrades the risk pool. Carriers respond by filing rate increases. This channel operates on a 3–9 month lag, which is consistent with the 6m correlation.

Second, shared supply-chain co-movement. Food and automotive parts inflation are both downstream of commodity-price and logistics-cost cycles. The Food@Home signal may be picking up the same supply-chain pressure that drives parts inflation, rather than an independent income-compression mechanism. If true, Food@Home is a redundant predictor once VehicParts is included — and its interaction term with VehicParts (interaction R² gain = +0.246 in the V2-C best triple) is consistent with redundancy that turns nonlinear under joint elevation. When both food and vehicle repair costs are simultaneously high, the joint effect on insurance rates exceeds the sum of independent effects.

The V2-C analysis does not resolve which mechanism is correct. Both are plausible. Either way, Food@Home at 6m survives pre-COVID stability and belongs in any multi-signal directional model.

Electricity@12m — the tenants insurance proxy chain

CPI Electricity at 12m lag has full-window ρ = 0.432, attenuating to 0.324 pre-COVID. It survives the stability threshold (delta = −24.8%), though attenuated rather than stable. The mechanism hypothesis: energy cost inflation feeds into rent, which feeds into landlord exposure, which feeds into tenants insurance policy pricing. This is the shortest plausible structural chain for the tenants-insurance CPI substitute used here — the Y variable is CUSR0000SEHG, a tenants/household series, not the retired auto insurance CPI. Electricity@12m may be more mechanistically relevant to this specific outcome than it would be to a pure auto-insurance model.

The interaction that matters most

The best triple from V2-C is not three independent directional signals. It is three predictors with large interaction gain: VehicParts@18m + Food@6m + Unemp@24m produces interaction R² of 0.512 vs main-effects R² of 0.308. The interaction gain of 0.205 is the largest found across all 455 three-way combinations tested.

The practical reading: these three pressures are compounding, not additive. Periods of simultaneous repair cost inflation, food price pressure, and elevated unemployment produce insurance rate acceleration that exceeds linear extrapolation. Rate buyers in environments with all three elevated should expect compounding rate action; environments with only one elevated are directionally predictive but not synergistic.

The kill-log

This is the most important section. The findings that did not survive are as informative as the ones that did.

Kill 1: CPI Used Cars/Trucks at 24m lag — eliminated

V1’s random forest flagged CPI Used Cars/Trucks at 24m lag as the fifth-ranked feature by permutation importance (0.017 ± 0.002). The mechanism hypothesis was that used-car price cycles lead insurance rates through the salvage-value and replacement-cost channels.

V2-C killed it. Full-window ρ = −0.025. Pre-COVID ρ = −0.099. Neither is meaningfully different from zero. The percentage change calculation in the V2-C log shows −301.8% — an artifact of dividing a near-zero full-window estimate by a slightly-different near-zero pre-COVID estimate. The practical reading: both estimates are noise.

The honest diagnosis: V1’s RF permutation importance picked up a confounded signal. CPI Used Cars/Trucks had extreme pandemic-era variance (the 2021 used-car price spike was unprecedented). The RF assigned importance based on the COVID episode. When you remove COVID from the stability test, the residual correlation is indistinguishable from zero. This is the standard failure mode for tree-based models trained on structural-break data.

Status: KILLED. Do not cite Used Cars/Trucks at 24m lag as a rate predictor. The V1 importance estimate was noise.

Kill 2: Unemployment Rate at 24m lag — weakened to explicit caveat

V1’s RF ranked Unemployment Rate at 24m lag fourth by importance (0.035 ± 0.004). Full-window ρ = 0.320. This was the weakest of the top-five pairwise signals but still positive and directional.

V2-C weakened it: pre-COVID ρ = 0.496, which is actually higher, not lower. The percentage change calculation shows 55.2% — meaning the full-window estimate is weaker than the pre-COVID estimate. This is the opposite of a COVID-inflation artifact; COVID appears to have muted the unemployment-rate correlation rather than inflating it. The most likely explanation: the 2020 unemployment spike was extreme but brief, and the insurance rate response in 2021–2022 was driven primarily by supply-chain parts costs rather than unemployment dynamics. The 2008–2009 recession, which falls within the pre-COVID window, had a more protracted unemployment cycle with a cleaner rate-response structure.

The Granger test adds a limitation: reported lag of 24 months exceeds the maximum tested horizon of 12 months in the bivariate Granger framework. The formal Granger result cannot speak to whether unemployment at 24m leads insurance rates in any causal sense.

Status: WEAKENED. Directional signal at full window (ρ = 0.320) is real but modest. Pre-COVID window is stronger (0.496) but the COVID episode complicates the mechanism interpretation. Use as a secondary directional indicator. Do not lead any rate-cycle claim with this signal. Explicit caveat required.

Kill 3: The 6-month consensus — downgraded, not killed

The six-month consensus is not wrong. V1 found positive ρ at six-month lag for CPI Motor Vehicle Parts (ρ = 0.421 pre-COVID; not displayed in the headline table above because the peak is at longer lags). The consensus is citing a real and directionally correct lag relationship.

What V1 and V2-C establish is that the six-month lag is not the peak. The peak is 12–18 months. The six-month reading is a shoulder of the correlation structure, not the top. Popular insurance commentary has been measuring at the shoulder and inferring the peak.

Status: DOWNGRADED, not killed. The consensus lead-time estimate is correct in direction, wrong in timing. The Rate Authority claim is specific: the peak is at 12–18 months, not 6. This is not a refutation of the consensus mechanism. It is a timing correction to the consensus lag estimate.

Kill 4: Granger causality — zero graduations at p < 0.01

V2-C ran bivariate Granger tests on the five top-ranked predictors at their reported lag structures. The test setup: ADF-differenced series where non-stationary, optimal VAR lag by AIC, F-test threshold p < 0.01.

Result: no predictor graduated. Zero of five.

The closest result was CPI Food at Home, which showed p = 0.023 at lag 7 and p = 0.035 at lag 8 in the full bivariate specification — but the reported causal lag for this predictor is 6 months, and p at lag 6 is 0.112. The result did not clear the rate-specific threshold.

There is a structural limitation in the test design: the strongest predictors in V1 and V2-C operate at 12–24 month lags. The standard bivariate Granger framework tested lags up to 12 months. The 18m and 24m predictors cannot be evaluated by a test that tops out at 12m. The honest reading is not “these predictors are Granger-non-causal” but rather “the test design cannot speak to lags beyond its window.” The test design is a limitation of scope, not evidence against the predictors.

Status: NO GRANGER GRADUATIONS at the reported lag structures and the standard 12m test window. The finding is a test-design limitation, not a structural null. Any publication using Rate Authority’s correlation findings should note: “Granger causality not confirmed at p < 0.01; formal causal claims are not supported by the current test.”

Kill 5: V1’s RF out-of-bag R² = 0.87 — downgraded

The V1 random forest on 120 features against CPI Tenants Insurance YoY changes produced an out-of-bag R² of 0.871. This number should not be cited as evidence of predictive performance.

The problem is three-fold. First, the 120 features include 24 predictors at five lags each — a highly collinear feature space that inflates apparent explanatory power even in out-of-bag estimation. Second, the full 2001–2026 window contains the 2020–2022 COVID supply-chain episode, which generates extreme co-movement across virtually all cost-inflation series. An RF trained on this window will assign high importance to any cost-inflation predictor that co-moved during COVID. Third, the RF permutation importance rankings are defensible; the OOB R² headline is not. Permutation importance is less sensitive to training-set regime outliers than are scalar R² estimates.

The pre-COVID linear model estimates tell a more honest story: the best triple (VehicParts@18m + Food@6m + Unemp@24m) achieves adj R² = 0.499 with interactions on the full window, and the best quad (Shelter@0m + VehicParts@24m + Electricity@12m + Food@6m) reaches adj R² = 0.593. These are substantial — half the variance in consumer insurance rate changes is potentially predictable from a four-predictor model — but they are not 0.87, and they have not been validated on pre-COVID-only fits for each combination.

Status: DOWNGRADED. The headline R² = 0.87 is contaminated by COVID supply-chain co-movement. Report as permutation-importance rankings, not scalar R². Best defensible model fits are 0.499–0.593 for top triples/quads on the full window.

What this does not claim

Eight gates — status as of 2026-05-22

The Rate Authority standard for a signal to graduate from directional_only to validated requires clearing all eight gates documented in the pre-registration piece. Current gate status:

Updated gate status (2026-05-22, end of day): 5 PASS / 1 FAIL / 2 PENDING. Gates 2, 4, 5, 6, 7b pass. Gate 1 fails on calibration precision specifically — test-window Brier 0.1281 against a 0.10 threshold, but the continuous out-of-sample Spearman is ρ=0.855 and the gate’s 0.20 kill threshold was not approached. The conviction-filtered subset (gate 5) collapses Brier to 0.0128 with 100% direction accuracy on n=30 high-conviction observations, indicating the model is well-calibrated when it is confident and over-confident on the middle of the distribution. This is a recalibration job (isotonic is the standard fix), not a signal-failure kill. Gate 3 needs a formal OLS recomputation; gate 8 awaits cycle-1 SHA-lock and 2028 forward resolution.

This piece publishes as confidence_tier: directional_only. That tier will not change until all eight gates clear. The V2-C pre-COVID stability results move this closer. They do not move it across the line.

No causal claim, no forward-return claim

The correlations documented here are pairwise or small-multivariate associations between lagged macro series and insurance rate proxies. Pre-COVID stability is evidence against a simple COVID-artifact explanation. It is not evidence of structural causation.

The mechanism chain — PPI shock → claims severity → rate filing → DOI approval → CPI print — is mechanistically plausible and documented in the framework piece. It is separately motivated. The statistical correlation and the mechanism narrative are two independent legs of the argument. Neither substitutes for the other.

No forward return or premium-change forecast appears in this piece. Quantitative forward claims will not be published until cycle-1 SHA-lock at 2026-07-15.

What is next

Gate 6 + 7b extension — both passed (2026-05-22)

The gate 6 + 7b extension at /research/gate_6_7b_validation_2026-05-22/ completed on 2026-05-22 with both gates clearing.

Gate 6 passed. Residualizing CPI Motor Vehicle Parts at lag-12 against the full confounder stack — month-of-year + CPI Shelter + CPI Medical + Unemployment@lag-24 + Federal Funds Rate — yields ρ=0.485 (n=279), barely changed from the raw-residualized 0.470. RMSE skill 0.165 against a predictor-free baseline. The primary kill mode — Unemployment@24m absorbing the Motor Vehicle Parts coefficient and exposing the apparent leading indicator as a recession-cycle alias — did NOT fire. UNRATE@24m independently fails gate 6 at ρ=−0.066, confirming it was a confound rather than the structural driver.

Gate 7b passed. The lag-12 relationship replicates on two of three alternative outcomes. CPI Transportation Commodities NEC (CUSR0000SETD, the closest available FRED proxy for the retired CPI Motor Vehicle Insurance) returns residualized ρ=0.548. CPI All-Items Less Food and Energy returns ρ=0.363. CPI Used Cars and Trucks (CUSR0000SETA02) correctly fails to replicate at ρ=0.00 — confirming the signal is specific to insurance and broad-inflation outcomes, not to vehicle-price cycles. That is the strongest possible specificity result: the predictor is not just leading whatever happens to inflate; it is leading the insurance-pricing pathway in particular.

8-gate status: 4 PASS (gates 3, 4, 6, 7b) / 0 FAIL / 4 PENDING (gates 1 Brier walkforward, 2 BSS, 5 conviction-filtered subset, 8 SHA-lock + 2028 forward resolution). The piece remains at confidence_tier: directional_only until the remaining four clear.

Alternative outcome candidates

CPI Motor Vehicle Insurance (CUSR0000SETC01) is retired from FRED and returned HTTP 400 during the V1 data pull. If BLS reconstructs the series under a new identifier, the auto-insurance-specific correlation should be re-run and compared to the tenants-insurance substitute. The current results are directly applicable to renters insurance rate cycles; auto applicability requires a separate run on a valid auto-insurance CPI series.

NAIC state-product premium history for 2018–2022 is the highest-value data acquisition. Only 2023 cross-section data is currently available (153 rows). A multi-year state panel would allow state-by-state leading-indicator validation, which would be substantially stronger evidence than the CPI aggregate proxy.

SHA-lock schedule

• Cycle 1 SHA-lock: 2026-07-15 (after gate 6/7b extension and V2-C integration)
• Cycle 2 SHA-lock: 2026-09-01 (after alternative-outcome replication)
• Forward forecast resolution: 2028-01-15 (January 2028 BLS CPI monthly release)

The forward prediction to be locked on 2026-07-15: year-over-year change in CUSR0000SEHG for the twelve monthly observations January 2027 through December 2027 will be positively correlated with year-over-year change in CUSR0000SETC for the twelve observations January 2026 through December 2026, at Spearman ρ ≥ +0.40 with p < 0.05, and the correlation will peak in the lag window [10, 14] months. These thresholds are placeholders until the cycle-1 lock; they may be adjusted based on gate 6/7b results before locking.

Methodology

Data: BLS FRED monthly series, 2001–2026. Primary outcome: CPI Tenants/Household Insurance (CUSR0000SEHG), year-over-year changes. Predictors: CPI Motor Vehicle Parts (CUSR0000SETC), CPI Food at Home (CUSR0000SAF11), CPI Electricity (CUSR0000SEHF01), Unemployment Rate (UNRATE), CPI Used Cars/Trucks (CUSR0000SETA02), CPI Shelter/OER (CUSR0000SAH1), Industrial Production Manufacturing (IPMAN), PPI Copper (WPU102501). All correlations residualized on month-of-year fixed effects. N = 279–303 full window; N = 204–228 pre-COVID window (2001–2019).

Methods (V1): Spearman and Pearson pairwise correlations at 0, 6, 12, 18, 24 month lags (120 features × 5 lags). Random Forest (300 trees, max depth 5, min leaf 5, permutation importance 30 repeats). Triple and quad scan: all C(15,3) = 455 combinations of top-15 RF features, including 2-way and 3-way interaction terms. OLS on top triples and quads.

Methods (V2-C): Pre-COVID stability test — rerun pairwise Spearman ρ on 2001–2019 subsample (n ≈ 204–228). Verdict thresholds: stable if |delta| < 50% and no sign flip; weakened if |delta| ≥ 50%; killed if near-zero in both windows. Bivariate Granger causality: ADF stationarity, difference if needed, optimal VAR lag by AIC, F-test p < 0.01 threshold, lags 1–12 months tested. Triple/quad scan on full window with interaction terms; V1 baseline benchmark.

Limitations: CPI Motor Vehicle Insurance (CUSR0000SETC01) retired; CUSR0000SEHG is the primary substitute and CUSR0000SETD is the alternative-outcome anchor. Findings directly applicable to renters/household insurance and to the closest available auto-adjacent CPI proxy; direct auto-insurance applicability awaits BLS reconstruction of the retired series. Granger test horizon (1–12m) cannot formally evaluate 18m and 24m predictors — this is a test-design limitation, not evidence against the mechanism. Gates 1 (Brier walkforward), 2 (BSS), 5 (conviction-filtered subset Brier), and 8 (SHA-lock + 2028 forward resolution) remain. The piece stays at confidence_tier: directional_only until those clear.

Related pieces: Insurance Price Leading Indicators — Framework · The 12-Month Lag Pre-Registration · DOI Filings as Leading Indicator · SEC 10-Q Carrier Disclosures · Methodology: Conviction Tier · Methodology: Rate Authority

Pre-registration scope: chorus-insurance/outputs/consumer_rate_leading_indicator_validation_scope.md. Cycle-1 SHA-lock target: 2026-07-15. Cycle-2 SHA-lock target: 2026-09-01. Forward resolution: 2028-01-15 (BLS January 2028 CPI release).

Rate Authority Editorial. 2026-05-22. Operated by PolicyChat. Citation: Rate Authority. “Cross-Signal Correlation Analysis: What Predicts US Consumer Insurance Rates (2026 Findings).” https://rateauthority.org/indicators/cross-signal-correlation-analysis-2026/. Last updated 2026-05-22.