# Scoring data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → across policy and claims systems
A claims adjuster at a mid-size US carrier pulls up a auto-claim file to authorize a $14,000 payout. The vehicle identification number (VIN) field is blank, the loss date is three weeks after the policy's cancellation date, and the claimant's address has two different ZIP codes across two systems. Nobody invented this scenario: it is a completely ordinary Tuesday in claims operations. Before actuaries can trust that file for pricing or reserving, someone has to score it.
This lesson walks through how carriers actually do that scoring.
Insurance runs on data collected from many hands: agents, third-party administrators (TPAs, external firms that process claims on an insurer's behalf), telematics devices, repair shops, medical providers, and policyholders themselves filling out forms. Each handoff introduces error risk.
Unlike a retailer with one point-of-sale system, a carrier typically stitches together:
Actuaries pricing a book of business or setting reserves depend on all of this lining up. A "data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → score" is the carrier's data office answering one question before release: is this dataset fit for the decision someone is about to make with it?
Most carrier data-governance frameworks (often aligned loosely with DAMA International's Data Management Body of Knowledge) converge on four dimensions. Definitions matter here because non-technical stakeholders often conflate them.
1. Completeness: Are required fields populated?
Example: a claims feed with 10,000 rows where 400 are missing VIN, and VIN is required for subrogation (recovering costs from an at-fault third party's insurer). That's a completeness gap.
2. Accuracy: Does the value reflect reality?
Example: a policy record shows a homeowner's roof age as 5 years when an inspection photo shows visible wear consistent with 20 years. The field is populated (complete) but wrong (inaccurate).
3. Timeliness: Is the data available when needed, and current?
Example: a claim's reserve estimate (the insurer's set-aside for expected payout) hasn't been updated in 90 days despite new medical bills arriving. Stale data distorts reserve adequacy.
4. Consistency: Do values agree across systems and over time?
Example: the claimant's date of birth is 03/14/1980 in the policy system and 04/13/1980 in the claims system. Same person, two systems, two truths.
Some frameworks add validity (does the value conform to an allowed format or range, like a state code that must be one of 50 US postal abbreviations) and uniqueness (no duplicate policy or claim IDs). For this lesson we stick to the four most commonly weighted in carrier scorecards.
Say the data office samples 1,000 auto claims records before releasing them to the actuarial reserving team. It checks four rules, one per dimension:
| Metric | Rule checked | Records failing | Pass rate |
|---|---|---|---|
| Completeness | VIN field populated | 40 | 96.0% |
| Accuracy | Loss date falls within policy effective dates | 25 | 97.5% |
| Timeliness | Reserve updated within 30 days of last claim activity | 80 | 92.0% |
| Consistency | Claimant DOB matches across policy and claims systems | 15 | 98.5% |
A simple composite data-quality score is the average of the four pass rates:
Composite score = (96.0 + 97.5 + 92.0 + 98.5) / 4 = 96.0%Many carriers instead use a weighted score, because timeliness failures on reserves matter more financially than a stale ZIP code. A plausible weighting scheme:
Weights: Completeness 20%, Accuracy 35%, Timeliness 30%, Consistency 15%
Weighted score = (96.0×0.20) + (97.5×0.35) + (92.0×0.30) + (98.5×0.15)
= 19.2 + 34.1 + 27.6 + 14.8
= 95.7%Carriers typically set a release threshold (commonly cited internal benchmarks range from 95% to 98% depending on the dataset's downstream use, these are illustrative, not universal standards). Below threshold, the file gets kicked back to source-system owners for remediation before actuaries touch it.
This is conceptually similar to a data quality firewall: an automated gate that blocks low-scoring batches from flowing downstream, used in various forms by large carriers and reinsurers managing high-volume claims feeds.
Here is roughly what an accuracy check (loss date within policy dates) looks like as a rule, the kind of logic that sits inside a carrier's data-quality pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète →:
import pandas as pd
def check_loss_date_accuracy(df):
# df has columns: loss_date, policy_effective_date, policy_expiry_date
valid = (df['loss_date'] >= df['policy_effective_date']) & \
(df['loss_date'] <= df['policy_expiry_date'])
pass_rate = valid.mean() * 100
return round(pass_rate, 1)
# check_loss_date_accuracy(claims_df) -> 97.5Production versions run at much larger scale and log every failing record for a remediation queue, but the logic is this direct.
Data qualityData qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → isn't just an internal efficiency question. It has regulatory teeth:
The practical effect: a "data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → score" isn't a nice-to-have dashboard, it's evidence a carrier may need to show an examiner.
Vérification des acquis
1. Why is data quality scoring especially challenging for insurance carriers compared to a retailer with a single point-of-sale system?
2. In the opening scenario, the claimant's address shows two different ZIP codes across two systems. This discrepancy is best understood as an example of what underlying data quality problem?
3. A carrier's data office asks 'is this dataset fit for the decision someone is about to make with it?' before releasing data. What does this framing imply about data quality scoring?
4. Select ALL correct answers about the sources of data feeding into a carrier's decision-making that are described in the lesson.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers describing why the claims file example (blank VIN, loss date after cancellation, mismatched ZIP codes) matters before actuaries use it for pricing or reserving.
Sélectionnez toutes les réponses correctes.
There's no single universal industry benchmark (be skeptical of anyone citing one precise number as gospel), but useful reference points, flagged as estimates:
The honest takeaway: benchmarks are contextual. A marketing dataset can tolerate more noise than a reserving dataset feeding actuarial signoffs.