# Vehicle Defect Early-Warning Feed: free sample (as of 2026-10-13)

- `signals_weekly_sample.csv`: 21,462 rows, parent-company level, weeks 2025-10-13 to 2026-10-05
- `recalls_sample.csv`: 1,854 vehicle recalls received since 2025-10-05

The full product adds make-level signals, history from 2014, complaint records, daily
updates and point-in-time first-seen timestamps.

## Methodology

Source: NHTSA Office of Defects Investigation (US public domain): consumer complaints API and the
official recalls flat file (FLAT_RCL_POST_2010), refreshed daily. Coverage: 16 makes (Ford, GM brands,
Stellantis brands, Tesla, Toyota/Lexus, Honda/Acura, Hyundai, Kia), model years 2015 onward, vehicle
recalls only (equipment recalls excluded).

Signal: complaints per entity x component x week, compared with the previous 52 weeks. The surprise
score is the tail probability of the observed count under a negative-binomial baseline (Poisson when
the baseline is not overdispersed), expressed on the z scale. The current unfinished week is never scored.

Point-in-time: every record carries the time this feed first saw it. History loaded during the initial
backfill is flagged knowledge_time_estimated=true; everything after is observed live.

Backtest: for clean weeks (no recall of the same make x component in the prior 26 weeks, to avoid
recalls causing complaints), how often a new recall of the same make x top-level component followed
within 26 weeks, for signal weeks vs all weeks. Most signals are NOT followed by a recall; the feed is
an early-warning / monitoring input, not a recall predictor.


## Backtest (signal week -> new recall of same make x component within 26 weeks)

Base rate (any clean week): 9.7%

| surprise_z >= | signals | hit rate | lift |
|---|---|---|---|
| 2.0 | 2,473 | 14.8% | 1.52x |
| 3.0 | 482 | 14.5% | 1.5x |
| 4.0 | 122 | 15.6% | 1.6x |

Most signals are not followed by a recall: use as an early-warning input, not a prediction.

## Data dictionary

### signals

- `level`: 'make' (e.g. FORD) or 'parent' (stable parent company id, e.g. GENERAL_MOTORS)
- `entity`: Make name or parent_id
- `ticker_asof_week`: Listed ticker of the parent during that week (parent level only; FCAU -> STLA on 2021-01-16)
- `component`: NHTSA component category, or ALL for every complaint of the entity
- `week_start`: Monday of the week (complaints grouped by date filed with NHTSA)
- `week_complete`: False for the current, unfinished week (never scored)
- `complaints`: Distinct complaints filed that week (a complaint naming two vehicles of one make counts once)
- `crash / fire / injuries / deaths`: Totals reported in those complaints
- `baseline_mean_52w / baseline_std_52w`: Mean and std of the 52 weeks BEFORE this week
- `zscore_52w`: Plain z-score (legacy; over-flags small counts, prefer surprise_z)
- `p_value`: P(at least this many complaints | baseline), negative binomial or Poisson
- `surprise_z`: p_value on the z scale (p=0.00135 -> 3.0). Calibrated: ~0.6% of quiet weeks score >= 2.5
- `persist_3w`: How many of the last 3 complete weeks had surprise_z >= 2
- `ratio_13w_52w`: Average of last 13 weeks / 52-week baseline (trend context)

### complaints

- `odi_number`: NHTSA complaint id
- `make / model / model_year`: Vehicle named in the complaint
- `parent_id / ticker_asof_filed`: Parent company and its ticker on the filing date
- `date_filed / date_incident`: ISO dates
- `components`: List of NHTSA component categories
- `crash / fire / injuries / deaths`: As reported by the consumer
- `knowledge_time`: When this feed first saw the complaint (UTC)
- `knowledge_time_estimated`: True for history loaded in the initial backfill (not observed live)
- `note`: VIN and free-text narrative are intentionally excluded

### recalls

- `campaign`: NHTSA recall campaign number (e.g. 23V838000)
- `make / model / model_year / component`: Vehicle and recalled component
- `potentially_affected`: Units potentially affected (campaign-level figure from NHTSA)
- `report_received_date`: Date NHTSA received the defect report
- `owner_notification_date / record_created_date`: As published by NHTSA
- `defect / consequence / remedy`: NHTSA summary text
- `do_not_drive / park_outside`: NHTSA safety advisories
- `mfr_part_numbers`: Manufacturer part numbers involved
- `knowledge_time / knowledge_time_estimated`: First seen by this feed / backfill flag

## License & source

Derived from NHTSA data (US public domain). Not affiliated with or endorsed by NHTSA.
Statistical indicators only; not investment advice.