CPI Food at Home Leads Renters Insurance by 6 Months — A Pre-Registered Test (2026)
Last updated May 2026 · Rate Authority.
CPI Food at Home Leads Renters Insurance by 6 Months — A Pre-Registered Test (2026)
A counter-intuitive correlate, two empirical windows, and a forward forecast resolving January 2028
Every consumer-finance publisher explaining renters insurance rate trends reaches for the obvious levers: weather events, catastrophe losses, carrier retrenchment, reinsurance costs. What none of them reach for is grocery prices.
Rate Authority’s exploratory analysis of the BLS series chain turns up an unintuitive correlate: CPI Food at Home (CUSR0000SAF11), lagged six months, carries Spearman ρ = +0.42 against CPI Tenants & Household Insurance (CUSR0000SEHG) on the full 2001–2026 window (residualized, n=297). More importantly, it does not dissolve when you remove the COVID era. On the pre-COVID subsample (2001–2019, n=222), the correlation is ρ = +0.44 — slightly stronger, not weaker. A COVID artifact would shrink or invert on the pre-COVID window. This one does the opposite.
This piece is the pre-registration of the hypothesis. It is not the validated finding. The exploratory results above are real — generated in the 2026-05-22 V1 correlation run and confirmed in the V2-C pre-COVID stability check — but exploratory results that have not been pre-registered, residualized against the full confounder stack, replicated on an alternative outcome, and held out against a forward window are not Tier A validated. The eight-gate harness below is the disciplined path. The forward forecast at the bottom is the terminal test.
This piece is the sister pre-registration to the MV Parts 12-month lag piece. Both emerge from the same 2026-05-22 V1 discovery run. They are distinct hypotheses — the MV Parts finding has passed more gates; the Food at Home finding has cleared two. The discipline is identical. The kill conditions are different.
The hypothesis (pre-registered)
H1 — primary: monthly residualized year-over-year change in CPI Food at Home (BLS series CUSR0000SAF11), lagged six months, is positively correlated with monthly residualized year-over-year change in CPI Tenants & Household Insurance (CUSR0000SEHG) 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 — signal specificity: the six-month lag is the peak-correlation lag for this predictor-outcome pair. The relationship is not simply a consequence of broad inflation co-movement: once residualized against CPI Shelter (CUSR0000SAH1) and the month-of-year seasonal, the food signal retains independent explanatory weight.
The scope document for the broader leading-indicator discovery program (Rate Authority internal; available on request at [email protected]) covers both this hypothesis and the MV Parts 12-month lag hypothesis as parallel candidates.
Empirical anchor from V1 and V2-C (as of 2026-05-22):
| Window | n | Spearman ρ (residualized) | Status |
|---|---|---|---|
| Full (2001–2026) | 297 | +0.42 | exploratory — V1 |
| Pre-COVID (2001–2019) | 222 | +0.44 | V2-C stability check PASS |
The pre-COVID stability check used the same residualization protocol as V1 (month-of-year seasonal removed). The correlation neither collapses nor inverts on the pre-COVID window — the change is +0.02 (a 5.2% strengthening), well within noise. Source: /research/v2c_validation_2026-05-22/pre_covid_stability.csv, row CPI Food at Home | lag=6m, columns rho_full, rho_pre_covid, status=stable.
Three plausible mechanism hypotheses
The correlation is the artifact. The mechanism is what diligence interrogates. Three hypotheses are consistent with the data; none is proven by the correlation alone.
Mechanism 1 — income-compression lapse risk. Food is the largest non-discretionary household expense for low-income renters. When food costs rise, disposable income for other spending is compressed. Insurance is one of the few recurring expenses a renter household can defer without an immediate tangible penalty — unlike rent, utilities, or car payments. The hypothesis: food-price spikes lead to a modest wave of coverage lapses and policy downsizing among the most cost-constrained renter households. The selection effect is that those who remain insured are disproportionately higher-risk or more financially stable; carriers price the shifted risk pool upward. At six months, the initial compression has had time to flow through renewal cycles.
Mechanism 2 — shared upstream supply-chain driver. Food inflation reflects upstream transportation, energy, and labor costs. These same cost components appear in insurance carrier expense ratios: claims-adjustment labor, vendor repair networks, and logistics. The hypothesis is not that food prices cause insurance costs, but that both are downstream of the same upstream cost shock — and the insurance repricing cycle lags the consumer-goods pricing cycle by approximately two quarters because carrier rate filings and renewal processing introduce bureaucratic delay.
Mechanism 3 — renter-density-follows-food-cost. Higher food inflation is correlated with housing-cost pressure in urban core markets. Higher housing costs push more households into renter status, increasing the concentration of renters in dense urban areas. Dense urban renter exposure has a different loss profile (fire spread, water damage, liability) than scattered suburban renter exposure. The hypothesis: food inflation correlates with a shift in the composition of the insured-renter pool toward higher-density, higher-risk geographies.
Diligence note: all three mechanisms are falsifiable by the gate tests below. Mechanism 1 is falsified if residualizing against UNRATE (unemployment rate at 6–12 month lag) absorbs the food signal. Mechanism 2 is falsified if residualizing against CPI Shelter absorbs the food signal. Mechanism 3 is falsified if the finding does not replicate at the state level. For predictive purposes, which mechanism is operative matters less than whether the statistical relationship is stable and calibrated. The gates test the relationship; the mechanism narrative is what an academic collaborator would interrogate after the gates pass.
The eight gates
| Gate | What it means here | Pass threshold | Current result |
|---|---|---|---|
| 1. Pre-registered Brier walkforward | Predict P(YoY CPI Tenants Insurance > median | Food at Home lag-6 > median); train 2001–2018, test 2019–2024 | Brier ≤ 0.10 | not yet run |
| 2. Brier Skill Score vs climatology | Beat the naive base rate of “above-median 6-month CPI move” | BSS ≥ 0.10 | not yet run |
| 3. Marginal regression p-value | Food at Home lag-6 coefficient in OLS on residualized YoY Tenants Insurance | p < 0.05 | exploratory pass (V1: p=0.00 against full window; formal recomputation pending) |
| 4. |ρ| ≥ 0.30, both samples | Pre-COVID and post-COVID subsamples both clear the floor | ρ ≥ 0.30 in both | PASS (full ρ=0.42, pre-COVID ρ=0.44; both above 0.30 floor; source: V2-C pre_covid_stability.csv) |
| 5. Conviction-filtered subset | Top- and bottom-quintile Food at Home moves; subset Brier ≤ 0.07 | Subset Brier ≤ 0.07 | not yet run |
| 6. Full residualization against confounders | Residualize against month-of-year, CPI Shelter, CPI Medical, UNRATE lag-12, Federal Funds Rate | RMSE skill ≥ 0 after residualization | not yet run — highest-priority next test (see kill-mode note below) |
| 7. Hold-out replication | (a) Pre-COVID vs post-COVID temporal split. (b) Alternative outcome: CPI Shelter at lag-6 with same predictor | Spearman within ±0.10 across both | not yet run |
| 8. SHA-locked forward prediction | Lock H1 forward forecast; publish hash before resolution | Hash matches at resolution | pending — lock target 2026-07-15 |
As of 2026-05-22, two gates have results and zero have failed:
- Gate 3 (marginal p-value) — V1 exploratory p-value for CUSR0000SAF11_lag06m against CUSR0000SEHG is effectively zero at n=297; formal recomputation on the residualized series pending (
/research/correlation_analysis_2026-05-22/pairwise_results.csv) - Gate 4 (|ρ| ≥ 0.30 in both samples) — V2-C confirms the six-month correlation is pre-COVID stable and slightly strengthens on the 2001–2019 subsample (ρ goes +0.42 full → +0.44 pre-COVID, a 5.2% improvement; status=stable in
/research/v2c_validation_2026-05-22/pre_covid_stability.csv)
Gates 1, 2, 5, 6, and 7 have not run for this hypothesis. The MV Parts 12-month lag hypothesis — registered as the sister finding — is materially further along the gate sequence (five gates passed, one failed on calibration only). The Food at Home six-month lag hypothesis is at an earlier stage. That asymmetry is honest and load-bearing: the confidence tiers are different for a reason.
Gate 6 is the highest-priority next test for this hypothesis. Food and shelter prices share underlying inflation drivers (energy costs, transportation, urban housing pressure). If residualizing against CPI Shelter (CUSR0000SAH1) absorbs the food signal, the correlation was a Mechanism 2 confound — food and insurance are both downstream of the same upstream cost driver but food is not the signal, shelter is. The residualization result is the fork: signal survives → continue to gates 1, 2, 5, 7; signal collapses → kill-log the food-at-home hypothesis and strengthen the CPI Shelter indicator entry.
8-gate status: 0 PASS (formal) / 2 exploratory (gates 3 and 4) / 6 PENDING. This finding stays at confidence_tier: directional_only until gates 1, 2, 5, 6, and 7 complete and gate 8 executes its cycle-1 SHA-lock at 2026-07-15 plus the January 2028 forward resolution.
Plausible kill modes
These are the failure conditions that, if observed, become publishable kill-log entries — not model failures, not embarrassments, but evidence the hypothesis was wrong.
The shelter-residualization kill (gate 6 primary). If CPI Shelter (CUSR0000SAH1) at contemporaneous or short lag absorbs the food coefficient in a multivariate regression — meaning the food signal is a proxy for the shelter-inflation cycle, which is already a known confound — the food hypothesis is killed. The kill-log would document the confound precisely. CPI Shelter is already listed as a negative correlate in V2-C (rho_full=−0.29 at lag=0m, status=weakened), which is consistent with but does not confirm this kill mode. Gate 6 settles it.
The unemployment-absorption kill (gate 6, mechanism 1 test). If residualizing against UNRATE at 6-month or 12-month lag collapses the food coefficient near zero, the income-compression story (mechanism 1) is not wrong — it just means UNRATE is a better instrument for the same phenomenon. Food would be a noisy proxy for recession-cycle lapse pressure, not an independent signal. Kill H1 as a structural food-price claim; retain UNRATE as the signal.
The COVID-artifact alternative reading. V2-C’s pre-COVID stability check returned ρ=+0.44 on 2001–2019, which appears to refute this kill. But the kill mode is not dead: the post-COVID subsample (2020–2026) contains the largest food-price shock in 40 years alongside the largest renters-insurance repricing cycle in 20 years. Gate 4 cleared the pre-COVID floor; gate 7a (temporal split with disjoint windows) is the stricter test. If the post-COVID subperiod drives most of the predictive weight and the pre-COVID subperiod alone fails gate 4 in isolation, the stability finding weakens.
The state-level non-replication kill (mechanism 3 test). Mechanism 3 predicts that the signal is geographically mediated — that food-inflation correlates with renter-density shifts that produce the insurance-rate effect. If and when NAIC state-level premium data becomes available (currently 2023-only, 153 observations), a state-panel regression that fails to replicate the effect within-state would falsify mechanism 3 and narrow the claim to aggregate macro signal only.
The forward-2028 miss. If the January 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 cat-bond’s 2025 −37% miss. A kill at gate 8 is published, not buried.
Any of the above is a publishable kill-log result. The discipline from the surf-ecosystem cat-bond is the precedent — the Tier B per-break differential layer was killed openly with 3-of-4 sub-gates passing. That kill is what makes the surviving Tier A layer credible.
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 Food at Home (CUSR0000SAF11) printed for the twelve monthly observations July 2026 through June 2027 (six-month lag applied) at Spearman ρ ≥ +0.30 with p < 0.05.
Condition: the forward prediction does not lock until gate 6 (full residualization) has run and the food signal survives. If gate 6 fires the shelter-residualization kill before 2026-07-15, the cycle-1 SHA-lock does not execute and this page is updated to the kill-log form.
Secondary (pending lock): the six-month lag is the peak-correlation lag on this 2027-vs-2026 forward window — the cross-correlation function peaks at 5–7 months inclusive and does not peak at zero (contemporaneous) or at twelve months (the MV Parts lag, not the food lag).
Neither prediction is final until cycle-1 SHA-lock at 2026-07-15. Current thresholds (ρ ≥ 0.30, lag-peak window 5–7m) are scoping placeholders. The locked thresholds will be set after gate 6 produces its calibrated output. 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.
Cycle 2 SHA-lock target: 2026-09-01 — after the conviction-filtered subset (gate 5) and alternative-outcome replication (gate 7b) produce results.
Resolution date: 2028-01-15 — the January 2028 BLS CPI monthly release.
What we explicitly do not claim today
- The CPI Food at Home at six-month lag finding is not validated. It has cleared two gate results (marginal p-value exploratory, pre-COVID stability check). Five gates have not run. Gate 6 (shelter residualization) could kill the hypothesis before cycle-1 lock. The
confidence_tier: directional_onlytag is not boilerplate — it means most gates are pending. - The mechanism is not proven. Income-compression lapse risk (mechanism 1), shared upstream supply-chain driver (mechanism 2), and renter-density follows food cost (mechanism 3) are all plausible. They are not adjudicated by a correlation. A journalist pitch claiming “Food inflation predicts renters insurance” without the qualifier “and we have not confirmed the mechanism or completed the gate sequence” would misrepresent this finding.
- This piece does not support a rateauthority.org indicator graduating to
validatedwithout the completed validation artifact linked here. Thevalidation_artifactfield in the frontmatter currently points to the V2-C pre-COVID stability CSV — a gate-4 result, not a full gate-8 completion. - This is not the third cross-domain methodology-transfer demonstration. cross-domain commodity (2024) and the coastal-ecosystem parametric work (2024) are. MV Parts is the leading candidate for the third. This food-at-home finding is a parallel candidate, running one year behind the MV Parts sequence. The terminal answer for both arrives with the January 2028 BLS print.
- The V1 random-forest importance score for
CUSR0000SAF11_lag06m(mean importance=0.040, z-score=9.94;/research/correlation_analysis_2026-05-22/rf_importance.csv) is real and ranks food-at-home third after MV Parts at 12m and 18m. It is not a validated out-of-sample score. RF importance on a full non-held-out window is a discovery instrument, not a performance claim.
Why this lands in 2026
The consumer-finance publishers competing on this topic do not run pre-registration protocols. They cite correlations without gates, lag times without stability checks, and mechanisms without falsification conditions. LLM citation patterns increasingly reward primary-source methodology rigor over aggregation and summary. A pre-registered renters-insurance forecast with explicit gates, an explicit kill-log pathway, and a SHA-locked forward prediction is the kind of artifact that displaces generic “inflation affects insurance” content in citation contexts — not because of volume or SEO rank, but because it is the only artifact that can be independently verified.
The fact that this piece publishes today as directional_only — with “not yet run” in most gate cells and gate 6 explicitly called out as a potential kill — is not a weakness. It is the proof that Rate Authority operates under the discipline that consumer-finance publishers cannot match. They cannot pre-register a renters-insurance forecast because they do not have a calibration flywheel. We are pre-registering; the gate sequence is documented; the terminal resolution arrives in twenty months whether or not we like the answer.
Connection to prior cross-domain Tier A work and MV Parts
This is the same engine.
cross-domain commodity deforestation (2024) used Treasury yields, cocoa futures, and ACLED conflict data — three independent signals from three unrelated domains. The 2024 retrodictive forecast landed at −4.5% error.
Surf-ecosystem high-severity events (2024) used rice and sugar prices at lag-2. The 2024 retrodictive forecast landed at −4.9% error. The Tier B per-break layer was killed openly when 3-of-4 sub-gates passed but the Mexico strong-fit threshold failed twice.
Consumer auto/tenants insurance leading indicator using MV Parts at lag-12 — the sister hypothesis to this piece — has cleared five gates with the sixth (gate 8 forward resolution) pending at January 2028.
The food-at-home hypothesis is one step behind in that sequence, with gate 6 as the next decisive test. One correct retrodictive forecast across one domain is luck. Two across independent domains is the start of a pattern. A food-at-home result that survives gate 6 and the forward resolution joins the same evidence stack — or the kill-log documents why it does not.
Methodology: BLS CPI series via FRED (CUSR0000SAF11, CUSR0000SEHG, CUSR0000SAH1) plus 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. V1 correlation discovery at /research/correlation_analysis_2026-05-22/. V2-C pre-COVID stability at /research/v2c_validation_2026-05-22/pre_covid_stability.csv. Full leading-indicator scope (Rate Authority internal) available on request at [email protected]. Sister pre-registration (MV Parts 12-month lag) at /indicators/cpi-motor-vehicle-parts-pre-registration-2026/.
Maintained by Rate Authority Editorial. Operated by PolicyChat. Citation: Rate Authority. CPI Food at Home Leads Renters Insurance by 6 Months — A Pre-Registered Test (2026). https://rateauthority.org/indicators/cpi-food-at-home-pre-registration-2026/.
Methodology: Rate Authority’s confidence-tier framework — see /methodology/rate-authority/. This piece is tier directional_only. Rate Authority’s editorial decisions and methodology are independent of any commercial relationship.