Vehicles on the register with no insurance, and renewals at risk
Why this matters to IIB. IIB's VAHAN work is about finding vehicles on the road with no live policy. This page sets policies in force against the real VAHAN register by state, and shows which policies may lapse in the next 60 days.
Vehicles on the VAHAN register, minus vehicles with a live policy. The register side is real. The policy side is a demo sample, scaled up to the whole market.
Why only IIB can do this. Only IIB sees policies from every insurer, so only IIB can do this match.
What this demo can show. Open VAHAN data gives counts, not vehicles, so this page estimates the gap by state. With IIB's own repository and VAHAN link, the same page would match each registration to a live policy.
Who acts on lapses. Each insurer would get only its own policies likely to lapse, so it can remind the owner before cover ends. IIB would keep the state totals for transport departments and awareness drives.
Only IIB sees a customer who moves insurer. A renewal with any insurer counts as a renewal. In the demo data 39,635 of 4,62,318 renewals moved to another insurer. A single insurer would count those as lapses.
The uninsured estimate is a little more than 33.51 crore minus 22.55 crore. In 17 state and vehicle class groups the insured estimate is above the register count. Each of those counts as zero, not as a negative.
1. The uninsured gap by state
Map of states
Darker means more uninsured vehicles. Two-letter codes are states and union territories, for example RJ is Rajasthan. Hover for the full name. Click a state to see its renewals at risk in part 2.
Ranked by uninsured vehicles
Rajasthan two-wheelers were built into the test data with a low insured share, set at 40%. The page finds 31.5%. Real figures would come from IIB data.
Cross-check. For 2022, VAHAN registrations over the same 15-year window add up to 28.6 crore. The MoRTH Road Transport Year Book counts 35.4 crore registered vehicles on 31 March 2022. That is 81%. So the register count is likely low, and the true uninsured gap is probably larger.
2. Policies likely to lapse in the next 60 days
The lapse forecast scores every demo policy that ends between 7 October 2026 and 5 December 2026. A policy counts as lapsed if it is not renewed with any insurer within 30 days of its end date. In the demo sample about 11,353 are expected to lapse. Scaled up the same way as the insured count, that is about 84 lakh renewals at risk across the market. Insurer is kept out of the model, so no insurer is scored up or down for being itself.
Renewals at risk by state, top 12 of 36 states and union territories
Expected lapses add up each policy's chance of lapsing. Policies already renewed count as zero. The lapse rate is expected lapses as a share of the policies not yet renewed. Click a state for its detail.
| State | Policies expiring | Already renewed | Expected lapses | Lapse rate |
|---|---|---|---|---|
| Uttar Pradesh | 7,615 | 305 | 1,636 | 22% |
| Maharashtra | 6,758 | 258 | 1,349 | 21% |
| Tamil Nadu | 3,956 | 149 | 888 | 23% |
| Gujarat | 3,705 | 146 | 761 | 21% |
| Karnataka | 3,525 | 143 | 739 | 22% |
| Madhya Pradesh | 3,415 | 129 | 719 | 22% |
| West Bengal | 2,372 | 87 | 497 | 22% |
| Rajasthan | 2,079 | 70 | 489 | 24% |
| Bihar | 1,945 | 76 | 447 | 24% |
| Andhra Pradesh | 2,037 | 86 | 441 | 23% |
| Kerala | 2,228 | 90 | 429 | 20% |
| Telangana | 2,122 | 74 | 413 | 20% |
One state in detail
Click a state on the map, in the ranked list or in the table to see its renewals at risk by vehicle class. Detail by RTO (Regional Transport Office) is for the Motor team.
How good is the forecast? Given one policy that lapsed and one that renewed, the model picks the lapsed one about 74 times in 100. It is checked each time it scores new renewals, to see if it still fits. All of this is on synthetic data. Real accuracy needs a pilot.
How the lapse forecast was checked
Model card
| Predicts | Not renewed within 30 days of expiry |
| Method | Gradient boosted decision trees (HistGradientBoostingClassifier), version lapse-hgb-1.0 |
| Learnt from | Policies that ended 1 October 2024 to 30 June 2026 (5,13,159) |
| Tested on | Policies that ended 1 July 2026 to 30 September 2026 (77,461), kept out of the learning |
| How well it ranks | AUC 0.738 on the test policies, 0.745 on the policies it learnt from. AUC is the chance that a lapsed policy scores higher than a renewed one. 0.5 is a coin toss and 1 is perfect. |
| Average error | Brier score 0.151. The average squared gap between the forecast chance and what happened. Lower is better. |
| What it looks at | Vehicle age, vehicle class, cover type, sales channel, premium change at renewal, past lapses, whether the RTO is in a city, state. |
| Left out | Insurer is kept out of the model, so no insurer is scored up or down for being itself. |
| Trained | 7 Oct 2026, 06:55, on data to 30 September 2026 |
Do forecast lapse rates match what happened?
The test policies are split into 10 groups of equal size, from the lowest forecast to the highest. Each point is one group's forecast lapse rate against the rate that happened. The closer the two lines, the better.
What drives a lapse most
Each input is shuffled in turn, and the bar shows how much the AUC falls. A longer bar means the input matters more.
Has the mix of vehicles changed since the model learnt?
For each input, the change score (PSI, population stability index) compares the policies the model learnt from with the policies it is scoring now. Above 0.2 the mix has changed enough to retrain.
| Input | Change score (PSI) | Status |
|---|---|---|
| Vehicle age | 0.117 | worth watching |
| Vehicle class | 0.005 | stable |
| Cover type | 0.009 | stable |
| Sales channel | 0.002 | stable |
| Premium change at renewal | 0.092 | stable |
| Past lapses | 0.000 | stable |
| Whether the RTO is in a city | 0.000 | stable |
| State | 0.001 | stable |
No input has changed enough to need retraining. This check runs each time the model scores new renewals.