What I did
The data
Three free files from NHTSA, the US road safety agency: owner complaints, safety recalls, and its defect investigations.
What gets ranked
Every vehicle and part that owners complained about in the past year, for example "2023 Honda CR-V, steering". That is about 22,000 combinations each month.
The method: logistic regression
For each vehicle and part I collect 73 facts known at the end of the month: complaint counts and growth, crash and fire reports, vehicle age, recent recalls and open investigations. Logistic regression learned a weight for each fact from 2015 to mid-2021, where the outcomes are known. It adds the weighted facts into one score: the estimated chance of a recall in the next six months.
The test
For each of 36 months, January 2023 to December 2025, I sort every vehicle and part by its score, take the top 100, and count how many were really recalled in the following six months.
Result: 17.3% of the logistic regression top 100 were recalled within six months
I ran the same test on three ways of choosing the 100.
The same test for six ways of ranking
Share of the monthly top 100 recalled within six months. The dot is the average over 36 months. The line is the margin of error.
What I found
- Logistic regression beat ranking by complaint count in all 36 months when the test was replayed month by month.
- Recall history is the stronger signal. A vehicle's and maker's past recalls predict the next one better than complaints do.
- Complaints matter most for vehicles with a clean record, where there is no recall trail to follow.
- Reading the complaint text did not help. It made the ranking slightly worse.
- Precision is modest. About of every 100 listed are not recalled within six months.
The watch list, month by month
Pick a month. The list is what the model would have produced at the end of that month, using only what was known then. The last column shows what happened afterwards.
| Rank | Vehicle | Part | Complaints, 12 months | Rank by complaint count | What happened next |
|---|
No rows match these filters.
Select a row to open that vehicle in the look-up tab.
Look up a vehicle
Covers the vehicle-and-part groups that reached the top 250, were recalled in 2023 to 2025, or had 10 or more complaints in a year.
Try one of the cases from the report: , , .
Complaints received per month
Dashed lines mark the month a recall was filed.
Rank on the monthly list
Rank 1 is the top of the list. Gaps are months when the group was not ranked, for example the 12 months after its own recall.
Recalls on this vehicle and part
Recent owner complaints
Replayed month by month, the list held up
For each month from January 2023 to December 2025 the model was trained only on outcomes already known, then ranked that month. This is how it would have run as a live system.
Recalled vehicles in the monthly list
How many of the listed groups were recalled within six months. Hit counts swing because one recall can cover many model years at once.
Where the lift comes from
Top-100 hit rate on 2023 to 2025 for models given different information.
How early the list warned
Months continuously on the top-250 list before the filing, for the recalled groups it caught (recalls filed 2024 and 2025).
Retraining and the held-back month
Open investigations and recall records are the warning signs that matter most
I rank the warning signs by what the model loses when that kind of information is taken away.
Top-100 hit rate when one kind of information is removed
Removing the share features or the complaint counts changes almost nothing. Shares of complaints by part were built and tested, and they added no measurable lift.
What pushes the risk up or down
The 14 largest weights in the final model. Right raises the risk, left lowers it. Overlapping complaint counts trade off against each other, so read them as a group.
Risk by part, other things equal
Hover to compare how often each part was recalled in the training years and in 2023 to 2025. The rates shifted, so part effects go stale.
Three cases
The results are modest, and I know where the gaps are
Limits
- Most listed vehicles are not recalled. The list is a way to decide where to look first.
- About 70% of recalls have no earlier complaint on the same vehicle and part, so no complaint-based model can reach them. The other 30% cover 94% of recalled vehicles.
- The list leans on a maker's recall record. Makers that recall often rank high even with few complaints.
- Part matching is coarse. Complaints and recalls are linked on 27 broad part groups.
- No sales data. Shares of complaints stand in for true failure rates.
- The 2023 to 2025 results were seen once before the final design. One later month was held back and never looked at until the end.
How it was tested
- Forward in time only. The model is trained on the past and scored on the future.
- No look into the future. Every input for a month uses only records dated up to that month.
- A pause after a recall. Complaints jump once a recall is announced, so a group leaves the list for 12 months after its own recall.
- Why not a random split?
Complaints build up before a recall, then spike
Average complaints per month for a recalled vehicle and part, lined up on the filing month.
Linking recalls to earlier complaints
Share of 2015 to 2025 recalls with at least one earlier complaint on the same vehicle and part.
Data sources
Three free files from the NHTSA datasets page, downloaded November 19, 2026: owner complaints, safety recalls since 2010, and defect investigations. The model is a logistic regression with 73 inputs. The comparison also covered random forest, gradient boosting and LightGBM.