ML to Predict Insurance Buyers

The Problem Hidden in Plain Sight

We analyzed over 50,000+ user journeys to answer one simple but critical question: "Which users are actually going to buy insurance — and when?"

At first glance, everything looked fine. Millions of users. High engagement. Strong traffic inflow. But one metric told a very different story.

Insurance conversions were disproportionately low — despite having users with relevant use cases, real-time triggers (expiry, RC checks), and clear monetisation opportunity.

This wasn't just a funnel problem. It was a system-level inefficiency.

Context

The Ecosystem We Were Operating In

The platform — a Vehicle Info ecosystem — served users who came with a clear, real-world intent: checking RC details, verifying ownership, reviewing insurance validity.

These are high-intent, real-world users — not passive browsers. Which makes them extremely valuable from a monetisation standpoint.

The Disconnect

Even with strong relevance — insurance modules were visible, timing seemed logical, traffic was high — conversion didn't scale proportionally.

Insurance modules were visible to the right audience
User intent (vehicle expiry, RC checks) was contextually relevant
Traffic volumes were substantial
Yet conversion rates remained stubbornly flat

Problem Statement

Problem Statement: We Were Treating Every User the Same

Every user saw the same insurance CTA. Every user received similar nudges. Every user experienced identical flows — regardless of where they actually were in their buying journey.

Core Issues

1. No Intent Differentiation

We couldn't distinguish between casual users, curious users, and ready-to-buy users. All three received identical treatment.

2. Generic User Experience

A user with expiring insurance and a user casually browsing received the same experience. No personalisation. No urgency calibration.

3. Lost High-Intent Users

Even ready-to-convert users faced unnecessary friction — they weren't being prioritised or fast-tracked to conversion.

We didn't have a traffic problem. We had an intent understanding problem.

Key Insight

Key Insight: Intent is Behavioral, Not Declared

We observed a critical pattern: users don't randomly convert — they exhibit predictable behavioural signals beforehand.

High-Intent Signals

Multiple searches in a session
Longer session duration
Insurance expiry within 30 days
Repeated engagement across days

Low-Intent Signals

Single interaction and exit
Quick page bounce
No CTA engagement
No return visits

Intent is not declared — it is inferred from behaviour. This was the breakthrough that changed everything.

If we can predict user intent using behavioural data, we can personalise their journey and increase conversions at scale.

Data Strategy

Data Strategy: 50,000+ User Sessions

We didn't guess. We built a data collection pipeline that captured granular behavioural signals across 50,000+ user sessions — structured, cleaned, and ready for modelling.

Behavioural Signals

Search count per session
Session duration (seconds)
Scroll depth (%)

Interaction Signals

CTA click events
Time on insurance page
Form abandonment rate

Contextual Signals

Days to insurance expiry
Vehicle age
RC verification frequency

Engagement Signals

Push notification interaction
Return visit frequency
Cross-session behaviour

Feature Vector:
[Search_Count, Session_Time, Expiry_Days, CTA_Clicks, Return_Frequency]

Derived Features:
→ Recency Score     (how recently did the user engage?)
→ Urgency Score     (how close is the expiry date?)
→ Engagement Index  (breadth + depth of interaction)
→ Intent Velocity   (speed of behaviour change over time)

ML Model

ML Model & System Architecture

Logistic Regression

Baseline model — fast to train, interpretable, good for initial validation

Gradient Boosting (XGBoost)

Production model — best for structured data, interpretable feature importance, scalable

The model output was a single Intent Score (0–100) per user, computed in near real-time and used to drive personalisation decisions downstream.

Intent Segmentation

0–30

Low Intent

Information-seeking. Focus on education and soft engagement.

30–70

Medium Intent

Curious but undecided. Focus on nudges, benefits, and comparisons.

70–100

High Intent

Ready to buy. Fast-track with minimal friction and strong CTA.

User Action → Data Capture → Feature Processing → ML Model → Intent Score → Decision Engine → Personalised Experience

Product Integration

Product Integration: Where Real Value Was Created

The ML model alone meant nothing. The value came from how the intent score was wired into the product experience.

🔥High Intent (70–100)

Fast-track journey with pre-filled forms
Minimal friction — no unnecessary steps
Strong, urgent CTA ('Renew Now – Expires in 3 days')
Priority placement in navigation and homepage

⚖️Medium Intent (30–70)

Nudge-based messaging (benefits, comparisons, social proof)
Contextual reminders tied to their vehicle data
Progressive disclosure — not overwhelming, not absent

❄️Low Intent (0–30)

Education-first content — build awareness, not urgency
Soft engagement: 'Did you know your insurance covers X?'
Retargeting via push notifications at better timing

Don't push harder — push smarter. The same message delivered at the wrong moment is just noise.

Results

Results: What Actually Moved

↑ 2–3×

Click-Through Rate

vs baseline CTR

+32%

Insurance Conversions

within 90 days

↓ Significantly

Drop-off Rate

high-intent segment

50,000+

Data Points Analyzed

user sessions

High-intent users converted at 2–3× baseline CTR
Revenue from insurance cross-sell grew significantly within 90 days of full rollout
Drop-off rates in the high-intent segment decreased as friction was reduced
Model confidence improved with each retraining cycle as more labelled data accumulated
ROI on paid retargeting improved as budget was reallocated to high-intent audiences

Not every user is equal. Growth comes from understanding intent — not just increasing traffic.

Learnings

Learnings & Trade-offs

Intent > Demographics

Traditional targeting by age, location, or device type was far less predictive than behavioural intent signals. A 45-year-old user who had checked insurance 3 times in a week was more valuable than a 25-year-old first-time visitor — regardless of demographic segment.

Timing > Placement

The right message at the wrong moment is just noise. A high-intent user who sees an insurance CTA at 11pm after their third session converts at dramatically higher rates than the same user seeing the same CTA on their first page view.

ML Needs Product Integration — Not Just a Score

An intent score sitting in a database is worthless. The real work was designing how the score influenced the product journey — which required deep collaboration between data science and product design.

Simplicity Wins in Production

The XGBoost model outperformed deep learning experiments not because it was more powerful — but because it was interpretable, debuggable, and reliable in production.

Cold Start Problem

New users have no behavioural history. We handled this via heuristics — defaulting to medium-intent treatment until enough data accumulated.

False Positives

Some users scored high but didn't convert. We tuned thresholds over multiple cycles and added contextual override rules.

Over-personalisation Risk

Making UX too aggressive for high-intent users created pressure and reduced trust. We balanced urgency with reassurance.

Future Scope

Real-Time Scoring

Move from near-real-time to sub-second intent scoring at the moment of each user action.

Deep Learning Models

Experiment with sequence models (LSTMs, Transformers) to capture longer behavioural patterns.

Cross-Platform Signals

Incorporate signals from web, app, and notification interactions into a unified intent model.

Reinforcement Learning

Let the system learn optimal messaging strategies directly from user response feedback loops.

How We Cut User Acquisition Cost by 30% Without Cutting Volume

Growth

How We Predicted High-LTV Users Using Behavioral ML (100K+ Sessions)

Machine Learning

How We Predicted High-LTV Users Using Behavioral ML (100K+ Sessions)

10 min read

Machine Learning

Building a Real-Time Intent Scoring Engine for Lead Conversion

9 min read

Machine Learning

Using ML to Reduce CAC by 31% via Smart Audience Segmentation

8 min read

All Case Studies

Browse all 30 breakdowns

Work With Me

Let's build something together

Connect on LinkedIn

Follow for more insights

Where to go next

Fintech Product Consultant
The consulting work this sits inside, with the rest of the evidence behind it.
Day 30 retention benchmark, fintech & banking
Published figures, traced to their sources — and where the sources disagree.
Day 1 retention benchmark, fintech & banking
Published figures, traced to their sources — and where the sources disagree.
Product Growth Score
Thirteen questions that locate which part of the funnel is actually constraining growth.

How We Used ML to Predict Insurance Buyers (50,000+ Data Points)

Results at a glance

The Problem Hidden in Plain Sight