E-Commerce Funnel & Conversion Analysis

The Goal

Understand where users drop off in the e-commerce funnel and build a predictive model to identify sessions likely to convert. By analyzing the customer journey from initial product view through cart abandonment to purchase, we aimed to uncover bottlenecks and develop actionable insights to improve conversion rates and reduce friction in the purchasing process.

The Data

Two complementary datasets provided a comprehensive view of customer behavior:

Retail Transactions

525K+ transaction records
Invoice data with quantities & prices
Customer IDs for traceability
Geographic distribution (countries)
Time-indexed for temporal analysis

E-Commerce Event Logs

1M+ event records (Oct-Nov 2019)
Event types: views, carts, purchases
Session-level granularity
User action sequences
Timestamps for behavior tracking

Data Preparation: Cleaned for duplicates, removed negative values, resolved type inconsistencies, and aligned schemas across datasets to ensure consistency and reliability throughout the analysis pipeline.

The Analysis

Funnel Analysis

Our initial analysis revealed a critical data quality issue that required methodological correction. The naive count of events showed impossible conversion rates, which led us to implement sequential funnel logic.

The first attempt used simple event counts, which produced a cart-to-purchase rate exceeding 100%—clearly impossible. This occurred because multiple purchase events could be attributed to the same session, and cart events weren't necessarily sequential predictors of purchases.

We corrected this by enforcing strict sequential funnel logic: each user session must follow the progression view → cart → purchase. This approach eliminated inflated conversion metrics and revealed the true customer journey:

Browse Sessions (View) 239,149

Add-to-Cart Sessions 10,115 (4.2%)

Purchase Sessions 5,751 (2.4% overall)

These metrics reveal significant drop-off points. The funnel analysis exposed that 95% of users never add items to cart, and of those who do, only 57% proceed to purchase. The visualizations below illustrate these critical conversion transitions:

Funnel conversion visualization showing view to cart to purchase stages

Funnel Stage Progression: Tracking user movement through browse, cart, and purchase stages

Conversion rate waterfall and drop-off analysis

Drop-off Analysis: Percentage of users progressing through each funnel stage

Predictive Model

To move beyond descriptive analytics, we built a logistic regression model to predict which sessions would result in a purchase based on observed behavior patterns.

The model uses three key session-level features:

Number of views: Quantity of products browsed in a session
Number of cart actions: Items added to cart (strongest signal)
Session length: Duration of user engagement

Model Performance Metrics:

Accuracy 93%

Area Under ROC Curve (AUC) 0.76

Precision 0.71

Recall 0.68

The model achieves 93% accuracy with a 0.76 AUC, indicating strong discriminative power. The logistic regression coefficients reveal the relative importance of each feature:

Add-to-Cart (Strongest): Each cart action increases the log-odds of purchase by ~0.67, translating to approximately 95% higher odds of converting. This is by far the dominant conversion signal.

Number of Views (Moderate): Each additional product view increases conversion odds by ~2.5%. While less impactful than cart actions, browsing behavior still contributes meaningful predictive signal.

Session Duration (Minimal): Session length showed negligible coefficient, suggesting that time spent doesn't significantly predict purchase intent—users know what they want quickly or not at all.

Key Findings

Browse-to-Cart Drop-off

95%

of users abandon before adding to cart

Overall Conversion Rate

2.4%

from initial view to purchase

Model AUC Score

0.76

strong predictive discrimination

Strongest Predictor

Cart Action

95% increase in purchase odds

Business Recommendations

Based on the analysis, we recommend prioritizing initiatives that address the largest friction points:

Cart Recovery Interventions Since 57% of cart sessions convert to purchases, capturing the remaining 43% via email reminders, incentives, or retargeting could yield significant revenue gains with minimal product changes.
Optimize Product Pages The 95% browse-to-cart drop-off suggests friction in the product discovery and decision process. A/B test page layouts, product information, and recommendations to reduce abandonment.
Personalize High-Intent Sessions Use the predictive model to identify high-likelihood converters in real-time and deliver personalized offers, reviews, or urgency messaging to accelerate purchase.
Monitor Funnel Metrics Weekly Establish dashboards tracking view-to-cart and cart-to-purchase conversion rates. Early detection of changes enables rapid intervention before revenue impact compounds.

What I'd Do Next

This analysis provides a strong foundation, but several extensions would deepen insights and improve actionability:

Compare Tree-Based Models: Evaluate random forests, XGBoost, or gradient boosting models to see if they capture non-linear relationships better than logistic regression. Feature importance plots would reveal additional behavioral signals.
Incorporate Product-Level Data: Analyze whether specific product categories, price points, or attributes have different conversion profiles. This enables targeted optimization beyond session-level interventions.
Segment New vs. Returning Users: Build separate models for first-time and repeat customers—their behavior patterns and conversion drivers are likely distinct, requiring tailored strategies.
Time-Series Analysis: Investigate temporal patterns (day of week, time of day, seasonal trends) to identify when users are most likely to convert and optimize marketing spend accordingly.
Cohort Analysis: Examine regional and demographic segments to uncover whether certain geographies or user cohorts have systematically different funnel performance, informing localization efforts.