# V2-C Validation Log — Insurance Rate Leading Indicator Analysis
**Generated**: 2026-05-22
**Based on V1 analysis**: `/Users/addieconner/policychat-content/research/correlation_analysis_2026-05-22/`
**V2-C scope**: Pre-COVID stability (2001–2019), triple/quad scan, Granger causality
**Y variable**: CPI Tenants/Household Insurance (CUSR0000SEHG) YoY changes, residualized on month-of-year
**Full window**: 2001–2026, n≈279–303 | **Pre-COVID window**: 2001–2019, n≈228

---

## Summary Box

| V1 Finding | Pre-COVID Status | Granger | Graduation |
|---|---|---|---|
| CPI Motor Vehicle Parts | lag=12m | stable (rho_pre=0.5398) | no_causality | — |
| CPI Motor Vehicle Parts | lag=18m | stable (rho_pre=0.5484) | no_causality | — |
| CPI Food at Home | lag=6m | stable (rho_pre=0.4435) | no_causality | — |
| Unemployment Rate | lag=24m | weakened (rho_pre=0.4961) | no_causality | — |
| CPI Used Cars/Trucks | lag=24m | weakened (rho_pre=-0.0986) | no_causality | — |
| Industrial Production Mfg | lag=12m | stable (rho_pre=0.2702) | not_tested | — |
| CPI Shelter/OER | lag=0m | weakened (rho_pre=-0.6398) | not_tested | — |
| CPI Motor Vehicle Parts | lag=24m | stable (rho_pre=0.5221) | no_causality | — |
| PPI Copper | lag=6m | stable (rho_pre=0.0266) | not_tested | — |
| CPI Electricity | lag=12m | stable (rho_pre=0.3244) | not_tested | — |

---

## Task 1: Pre-COVID Stability Results (2001–2019)

Pre-COVID subsample: n≈228 monthly observations (January 2001 – December 2019). This window excludes the 2020–2022 COVID supply-chain shock and 2023+ normalization period. Because V1 ran on 2001–2026 (n≈279–303), this subsample retains ~60% of the data while removing the most structurally distorted period.

### Stable Findings (effect size holds pre-COVID, Δ < 50%, no sign flip)

**CPI Motor Vehicle Parts | lag=12m** (lag=12m)
- Full-window rho: 0.4703 (n=291)
- Pre-COVID rho: 0.5398 (n=216)
- Delta: 0.0695 (14.8% change)
- Verdict: STABLE — correlation persists outside COVID distortion window

**CPI Motor Vehicle Parts | lag=18m** (lag=18m)
- Full-window rho: 0.5014 (n=285)
- Pre-COVID rho: 0.5484 (n=210)
- Delta: 0.047 (9.4% change)
- Verdict: STABLE — correlation persists outside COVID distortion window

**CPI Food at Home | lag=6m** (lag=6m)
- Full-window rho: 0.4215 (n=296)
- Pre-COVID rho: 0.4435 (n=222)
- Delta: 0.0221 (5.2% change)
- Verdict: STABLE — correlation persists outside COVID distortion window

**Industrial Production Mfg | lag=12m** (lag=12m)
- Full-window rho: 0.2579 (n=291)
- Pre-COVID rho: 0.2702 (n=216)
- Delta: 0.0123 (4.8% change)
- Verdict: STABLE — correlation persists outside COVID distortion window

**CPI Motor Vehicle Parts | lag=24m** (lag=24m)
- Full-window rho: 0.5082 (n=279)
- Pre-COVID rho: 0.5221 (n=204)
- Delta: 0.0139 (2.7% change)
- Verdict: STABLE — correlation persists outside COVID distortion window

**PPI Copper | lag=6m** (lag=6m)
- Full-window rho: 0.0247 (n=297)
- Pre-COVID rho: 0.0266 (n=222)
- Delta: 0.0019 (7.6% change)
- Verdict: STABLE — correlation persists outside COVID distortion window

**CPI Electricity | lag=12m** (lag=12m)
- Full-window rho: 0.4315 (n=291)
- Pre-COVID rho: 0.3244 (n=216)
- Delta: -0.1071 (-24.8% change)
- Verdict: STABLE — correlation persists outside COVID distortion window


### Weakened Findings (effect size drops >50% pre-COVID)

**Unemployment Rate | lag=24m** (lag=24m)
- Full-window rho: 0.3196 (n=279)
- Pre-COVID rho: 0.4961 (n=204)
- Delta: 0.1765 (55.2% change)
- Verdict: WEAKENED — COVID window inflated V1 estimate. Downgrade to directional_only with caveat.

**CPI Used Cars/Trucks | lag=24m** (lag=24m)
- Full-window rho: -0.0245 (n=279)
- Pre-COVID rho: -0.0986 (n=204)
- Delta: -0.0741 (-301.8% change)
- Verdict: WEAKENED — COVID window inflated V1 estimate. Downgrade to directional_only with caveat.

**CPI Shelter/OER | lag=0m** (lag=0m)
- Full-window rho: -0.2927 (n=303)
- Pre-COVID rho: -0.6398 (n=228)
- Delta: -0.3471 (-118.6% change)
- Verdict: WEAKENED — COVID window inflated V1 estimate. Downgrade to directional_only with caveat.


### Unstable Findings (sign flip or near-zero pre-COVID)

_None of the top-10 findings showed a sign flip pre-COVID._


### Insufficient Pre-COVID Sample

_All top-10 findings had n≥30 in the pre-COVID window._

---

## Task 2: Triple/Quad Scan

**Triples tested**: All C(15,3)=455 combinations of top-15 RF features, full sample (2001–2026).
**Baseline**: V1 best triple [Food@6m + Shelter@0m + Electricity@12m] → R²=0.4519, adj_R²=0.4512

### Top Triple (Full Sample)

**VehicParts@18m + Food@6m + Unemp@24m**
- adj_R² (with interactions): 0.4994
- R² (main effects only): 0.3075
- R² (with interactions): 0.512
- Interaction gain: 0.2045
- n: 278

### Top 5 Triples by adj-R² (interactions included)

| Rank | Features | n | R²(main) | adj_R²(int) | int_gain | Beats V1? |
|---|---|---|---|---|---|---|
| 1 | VehicParts@18m + Food@6m + Unemp@24m | 278 | 0.3075 | 0.4994 | 0.2045 | YES |
| 2 | Food@6m + Unemp@24m + Electricity@12m | 278 | 0.2807 | 0.499 | 0.2309 | YES |
| 3 | VehicParts@18m + UsedCars@24m + Food@12m | 279 | 0.4405 | 0.4985 | 0.0706 | YES |
| 4 | VehicParts@24m + Food@12m + NewVehicles@24m | 279 | 0.4614 | 0.4981 | 0.0493 | YES |
| 5 | VehicParts@18m + Food@6m + UsedCars@24m | 278 | 0.4473 | 0.4965 | 0.062 | YES |

### Top Quad

**Shelter@0m + VehicParts@24m + Electricity@12m + Food@6m**
- adj_R² (main effects): 0.5926
- R² gain over base triple: 0.1546


**Caveat**: All triple/quad R² values are full-sample (2001–2026). COVID episode inflates R² of any cost-inflation combination. These models should be treated as exploratory scaffolding, not publishable until pre-COVID sub-sample validation is run on each finalist triple.

---

## Task 3: Granger Causality Results

**Setup**: Bivariate Granger tests, lags 1–12m, ADF → difference if non-stationary, optimal lag by VAR-AIC, F-test p<0.01 threshold. Target = CPI Tenants Insurance (CUSR0000SEHG).

**Critical reminder**: Granger causality (predictive precedence) ≠ structural causation. Report as "granger_causal" only when p<0.01 AND pre-COVID stability check also passes.

**CPI Motor Vehicle Parts** (reported lag: 12m)
- ADF: X=differenced→stationary (p=0.0465), Y=differenced→stationary (p=0.2087) (differenced for stationarity)
- n (Granger): 302
- Optimal lag (AIC): 12m, F=1.223, p=0.266873
- All lags: lag1:p=0.172; lag2:p=0.095; lag3:p=0.177; lag4:p=0.286; lag5:p=0.413; lag6:p=0.498; lag7:p=0.594; lag8:p=0.559; lag9:p=0.176; lag10:p=0.152; lag11:p=0.196; lag12:p=0.267
- Note: p_at_reported_lag_12m=0.2668730495000067
- **Conclusion**: NO_CAUSALITY

**CPI Motor Vehicle Parts (18m)** (reported lag: 18m)
- ADF: X=differenced→stationary (p=0.0465), Y=differenced→stationary (p=0.2087) (differenced for stationarity)
- n (Granger): 302
- Optimal lag (AIC): 12m, F=1.223, p=0.266873
- All lags: lag1:p=0.172; lag2:p=0.095; lag3:p=0.177; lag4:p=0.286; lag5:p=0.413; lag6:p=0.498; lag7:p=0.594; lag8:p=0.559; lag9:p=0.176; lag10:p=0.152; lag11:p=0.196; lag12:p=0.267
- Note: reported_lag_18m_exceeds_max_tested_12m
- **Conclusion**: NO_CAUSALITY

**CPI Food at Home** (reported lag: 6m)
- ADF: X=differenced→stationary (p=0.0135), Y=differenced→stationary (p=0.2087) (differenced for stationarity)
- n (Granger): 302
- Optimal lag (AIC): 12m, F=1.7416, p=0.058339
- All lags: lag1:p=0.674; lag2:p=0.280; lag3:p=0.168; lag4:p=0.278; lag5:p=0.089; lag6:p=0.112; lag7:p=0.035; lag8:p=0.023; lag9:p=0.049; lag10:p=0.095; lag11:p=0.014; lag12:p=0.058
- Note: p_at_reported_lag_6m=0.11236717979153046
- **Conclusion**: NO_CAUSALITY

**Unemployment Rate** (reported lag: 24m)
- ADF: X=differenced→stationary (p=0.0025), Y=differenced→stationary (p=0.2087) (differenced for stationarity)
- n (Granger): 302
- Optimal lag (AIC): 12m, F=0.6585, p=0.790363
- All lags: lag1:p=0.476; lag2:p=0.476; lag3:p=0.605; lag4:p=0.232; lag5:p=0.340; lag6:p=0.397; lag7:p=0.566; lag8:p=0.598; lag9:p=0.599; lag10:p=0.729; lag11:p=0.827; lag12:p=0.790
- Note: reported_lag_24m_exceeds_max_tested_12m
- **Conclusion**: NO_CAUSALITY

**CPI Used Cars/Trucks** (reported lag: 24m)
- ADF: X=differenced→stationary (p=0.1149), Y=differenced→stationary (p=0.2087) (differenced for stationarity)
- n (Granger): 302
- Optimal lag (AIC): 12m, F=0.5518, p=0.879095
- All lags: lag1:p=0.985; lag2:p=0.404; lag3:p=0.431; lag4:p=0.572; lag5:p=0.110; lag6:p=0.176; lag7:p=0.245; lag8:p=0.332; lag9:p=0.379; lag10:p=0.448; lag11:p=0.561; lag12:p=0.879
- Note: reported_lag_24m_exceeds_max_tested_12m
- **Conclusion**: NO_CAUSALITY

---

## Task 4: Validation Graduation Assessment

### GRADUATION CANDIDATES (Pre-COVID stable + Granger-causal at p<0.01)

**None of the top-5 pairs simultaneously pass both gates.**

This is an honest null finding. The combination of pre-COVID stability and formal Granger causality is a high bar. Findings that fail one gate remain useful for directional modeling but should not be cited as validated causal indicators in external research.

### DIRECTIONAL_ONLY (pre-COVID stable but Granger NOT confirmed, or Granger not tested)

- **CPI Motor Vehicle Parts | lag=12m** (lag=12m): pre-COVID stable (rho_pre=0.5398), Granger=no_causality. Use as directional signal, not causal lead.
- **CPI Motor Vehicle Parts | lag=18m** (lag=18m): pre-COVID stable (rho_pre=0.5484), Granger=no_causality. Use as directional signal, not causal lead.
- **CPI Food at Home | lag=6m** (lag=6m): pre-COVID stable (rho_pre=0.4435), Granger=no_causality. Use as directional signal, not causal lead.
- **Industrial Production Mfg | lag=12m** (lag=12m): pre-COVID stable (rho_pre=0.2702), Granger=not_tested. Use as directional signal, not causal lead.
- **CPI Motor Vehicle Parts | lag=24m** (lag=24m): pre-COVID stable (rho_pre=0.5221), Granger=no_causality. Use as directional signal, not causal lead.
- **PPI Copper | lag=6m** (lag=6m): pre-COVID stable (rho_pre=0.0266), Granger=not_tested. Use as directional signal, not causal lead.
- **CPI Electricity | lag=12m** (lag=12m): pre-COVID stable (rho_pre=0.3244), Granger=not_tested. Use as directional signal, not causal lead.

### DOWNGRADED (weakened or unstable pre-COVID)

- **Unemployment Rate | lag=24m** (lag=24m): Full rho=0.3196 → Pre-COVID rho=0.4961 (Δ=0.1765, 55.2%). COVID window inflated V1 estimate. DIRECTIONAL_ONLY with explicit caveat: 'effect may be partially attributable to COVID supply-chain episode'.
- **CPI Used Cars/Trucks | lag=24m** (lag=24m): Full rho=-0.0245 → Pre-COVID rho=-0.0986 (Δ=-0.0741, -301.8%). COVID window inflated V1 estimate. DIRECTIONAL_ONLY with explicit caveat: 'effect may be partially attributable to COVID supply-chain episode'.
- **CPI Shelter/OER | lag=0m** (lag=0m): Full rho=-0.2927 → Pre-COVID rho=-0.6398 (Δ=-0.3471, -118.6%). COVID window inflated V1 estimate. DIRECTIONAL_ONLY with explicit caveat: 'effect may be partially attributable to COVID supply-chain episode'.

### UNCERTAIN (insufficient pre-COVID n)


---

## Honest Framing and Limitations

1. **Sample size bounds**: Full window n=279–303 (2001–2026); pre-COVID n≈228 (2001–2019). These are bounded but sufficient for pairwise correlations. Triple/quad models approach overfitting risk even on full window; all model R² estimates should be treated as upper bounds.

2. **COVID structural break is real**: Any finding where pre-COVID rho drops substantially vs full-window should be cited as "directional in normal rate cycles; amplified during COVID supply-chain periods." The 2020–2022 episode is not a random outlier — it is a structurally different inflationary regime.

3. **Granger ≠ causation**: Granger causality establishes predictive precedence (X t-1 improves forecast of Y_t). The structural mechanism (repair cost → loss trend → rate filing → CPI print) is separately documented and mechanistically plausible. The combination of Granger + stable pre-COVID rho + documented mechanism is the strongest defensible claim available with public data.

4. **Y series substitution**: The intended Y (CPI Motor Vehicle Insurance, CUSR0000SETC01) is retired/unavailable on FRED. CPI Tenants/Household Insurance (CUSR0000SEHG) is the substitute. Findings are directly applicable to renters insurance rate cycles; validity for auto insurance requires separate validation with a working auto-insurance CPI series (e.g., BLS microdata request, or use of NAIC state-level premium history when acquired).

5. **Novel vs rediscovered**: The 12-month lag (not 6m) on Motor Vehicle Parts is potentially novel relative to consensus actuarial timelines. Pre-COVID confirmation of this lag structure strengthens the novelty claim. Report as: "consistent with a state-filing-approval pipeline extension of ~6 months beyond the commonly assumed repair-cost-to-rate lag."

---

*Generated by V2-C validation agent, 2026-05-22. All correlations residualized on month-of-year fixed effects.*