Company
🏦 Santander
Skill Area
Product & Development
Tool Stack
BigQueryGoogle AnalyticsGoogle Tag ManagerPython
Overview
📉 Challenge | • At Santander, despite top of funnel optimizations on lead gen channels like PPC, SEO and website CR% optimization, low acceptance rates were costing valuable leads in terms of full funnel conversion (loan contract booking), as the existing risk model was trained on outdated channel data that didn’t align with the newer digital audience.
• Difficulty to scale without full funnel results despite quality lead generation at top of funnel |
💬 Opportunity | Extensive online user behavior data was available (from online marketing activities) that could be used to train a machine learning model, and potentially identify statistically significant patterns in customer actions that correlated with outcomes like missed repayments, defaults, and early settlements. |
🚀 Solution | After ensuring data validation and availability of in usable formats, analysis was performed to see which metrics and dimensions could possibly be relevant indicators for “Risk”.
Once done, an advanced machine learning algorithm was developed in Python to train a new risk predictive model that was eventually added in the middle layer of backend (where submitted leads went) and the main existing engine. This algorithm provided an ‘online scoring’ for risk that was consumed as another data point in the Risk assessment. |
Scope & Responsibilities
- BigQuery Integration: Set up BigQuery to capture hit-level data, creating a custom data pipeline for feeding behavior metrics into the risk team’s algorithm
- Improve data collection: Enabled additional behavioral dimensions data capture that could improve or be relevant for risk assessments, setting up additional GTM-based event tracking as required
- Compliance and stakeholder coordination: Besides closing collaborating with Risk on data topics, ensured legal and data compliance by collaborating closely with Legal team and Chief Data Officer (CDO).
Process
Results highlights
Accept rate | First stage accept rate increased by 33.5% (relative) after implementation of the new risk scoring algorithm, improving full funnel conversion by ~14% (relative) |
Lower loan default | Loan defaults decreased over next 2 years observed results from online source of leads |
ML model performance | 0.72 AUC-ROC score, i.e. 72% probability of correctly ranking a customer that would not default higher than one who would |
Other | Presented and won Digital Marketing Project of the Year award 2021 at Global Digital Marketing conference in Santander Group |