Master’s thesis: Fraud detection in animal insurance

Job Description

Are you interested in using data science, machine learning or statistics to tackle real-world challenges? Join our Claims Automation team for a master’s thesis exploring how to detect suspicious animal insurance claims, identify document anomalies and reduce incorrect reimbursements. We’ll shape the project together around your interests, available data and academic requirements.

More about the opportunity

Help us reduce claim-cost leakage and improve claims automation. Lower leakage from incorrect reimbursements, duplicate claims, errors, and potential fraud helps us keep insurance premiums competitive for our customers.
You will join the Claims Automation team who work on automation and efficiency in the claims domain. We are looking for a student interested in data science, machine learning or statistics.

Possible thesis directions:

• Estimate claim-cost leakage (Estimate the cost of undetected fraud and incorrectly reimbursed claims.)
• Detect suspicious claims (Identify claims that should be prioritized for additional manual review.)
• Detect document anomalies (Analyze invoices and receipts for duplicates, inconsistencies, or unusual patterns.)

Estimating Claim-cost leakage

Challenge

Known fraud and incorrect payments represent only the cases we have detected. The true cost of claim leakage is likely higher.

Objective

Develop a method for estimating lower and plausible upper bounds of claim-cost leakage, including potentially undetected fraud and incorrectly reimbursed claims.

Possible methods:

• Statistical estimation and uncertainty analysis
• Positive–unlabeled learning
• Audit- or sampling-based estimation
• Analysis of historical claim and investigation outcomes

Expected outcome

An estimate of hidden claim-cost leakage, including assumptions, uncertainty, and possible breakdowns by claim type, customer group, or other relevant segments.

Suspicious claim detection

Challenge

Most claims are legitimate, but some contain unusual patterns that may justify additional review.

Objective

Develop and evaluate a model that identifies and prioritizes unusual or potentially suspicious claims.

Possible signals

• Unusual claim frequency or claim amounts
• Claims shortly after policy inception
• Repeated or escalating claim patterns
• Patterns that differ from comparable customers or animals
• Claims close to deductibles, limits, or other thresholds

Possible methods

• Anomaly detection
• Statistical peer-group comparison
• Supervised or semi-supervised machine learning
• Explainable risk scoring

Expected outcome

A risk score or ranked review queue to support claims handlers in deciding which claims should receive additional manual review.

Anomaly detection of documents

Challenge

Claims are often supported by invoices and receipts that may be unclear, incomplete, duplicated, inconsistent, or unusual. We already use OCR in production to extract information from documents. There is an opportunity to build on this information to strengthen automated controls.

Objective

Investigate how extracted document data and document similarity analysis can identify anomalies and improve automated claims handling.

Possible areas

• Detection of exact and near-duplicate invoices and receipts
• Use of existing OCR-extracted fields, such as dates, amounts, clinic/supplier, invoice number, and line items
• Detection of document tampering and AI-generated documents
• Identification of documents that are not suitable for automated processing

Expected outcome

A prototype or framework for flagging documents for manual review due to inconsistencies, anomalies or potential fraud.