CLTV-driven customer acquisition strategy

A fintech mobile app was struggling with inefficient customer acquisition. They were spending heavily on broad targeting across social media platforms, but had no systematic way to identify and acquire high-value customers. The acquisition team was optimizing for volume metrics (clicks, conversions) rather than customer quality.

The core problem was a lack of connection between their acquisition strategy and long-term customer value. They had no way to predict which prospects would become valuable customers, leading to acquisition of low-value segments while missing high-value opportunities.

They needed a systematic approach to identify their most valuable customer segments, predict customer lifetime value for prospects, and use these insights to optimize their acquisition targeting and budget allocation.

I developed a two-part solution: a customer lifetime value prediction model and a lookalike audience optimization system. The CLTV model used the BG-NBD (Beta-Geometric/Negative Binomial Distribution) framework to predict customer behavior patterns - specifically purchase frequency and churn probability.

The model incorporated both transactional data (purchase history, order values, frequency) and behavioral features (engagement metrics, product preferences, seasonal patterns). For new prospects, I created a simplified CLTV scoring system based on observable characteristics during the acquisition funnel.

The second component was a systematic lookalike audience testing framework on Instagram and Facebook. Instead of creating audiences based on all customers, I segmented by predicted CLTV and created separate lookalike audiences for high, medium, and low-value customer segments. The system automatically tested different audience sizes (1%, 3%, 5%) and optimized budget allocation toward the highest-performing segments.

Implementation took 12 weeks. Weeks 1-3 focused on data collection and CLTV model development, analyzing 18 months of customer transaction data to identify behavioral patterns. Weeks 4-6 were model validation and segmentation, testing the BG-NBD predictions against actual customer behavior.

Weeks 7-9 involved setting up the lookalike testing framework on Meta platforms, creating audiences for each CLTV segment and implementing tracking. Weeks 10-12 were optimization and scaling, refining the model based on acquisition results and expanding to additional platforms.

The biggest technical challenge was handling the cold-start problem for prospects with limited behavioral data. I solved this by creating a simplified scoring model based on acquisition channel, initial engagement signals, and demographic proxies that correlated with high-CLTV customers.