ML
Sales & Marketing
Predicting Customer Lifetime Value (LTV) for ROAS Improvement
Our client operates a set of mobile applications, offering a variety of subscription plans. They sought to develop a predictive model that could accurately forecast Customer Lifetime Value (LTV) based on specific customer profiles and behavioral parameters for the purpose of making timely adjustments to advertising targeting to improve ROAS. Previously, the client used manual statistical calculations that would provide decent efficiency but caused a months-long lag in reaction time (until enough data was collected to calculate the actual user’s LTV). The goal was to build a model that would forecast the users’ potential LTV upon signing up and allow instantaneous adjustments to the online advertising strategy.

The Main Challenges Were

The project faced several significant challenges:

  • Lag in reaction time. In order to make ad targeting adjustments, the company needed to first collect statistically sufficient data, which could take up to 6 months, resulting in the advertising budget being partially spent on acquiring low-value users before proper adjustments could be made.
  • Unclear user segmentation, causing poor targeting of the most valuable users. The manual approach lacked precision in user segmentation, which affected the client's ability to properly distinguish user groups.
  • Inability to forecast LTV of statistically underrepresented user segments. The manual approach only allowed the client to look for patterns in user segments that were already present in the past data, which means some of the more valuable user segments could be overlooked.

Machine Learning-Driven Solutions

To address the challenges and meet the project objectives, we adopted a strategic approach focused on scalability, maximum data use and business value. The key initiatives included:

  • Immediate LTV Forecasting: Enables faster targeting adjustment by predicting customer value at the moment of sign-up.

  • Clear User Segmentation: Enhances targeting accuracy and marketing effectiveness by leveraging detailed profile and behavioral data.

  • LTV Forecasting for Underrepresented Segments: Expands market reach and ensures informed resource allocation by accurately predicting value for less common customer groups.

  • Solution Scalability: Future-proofs the system by allowing easy integration of new data types and sources with minimal disruption or retraining.

  • Seamless Model Updates: Ensures continuous operation and minimal downtime by enabling smooth transitions to new models without interrupting existing processes.

Results

The implementation of these solutions led to several key outcomes:

  • The solution that we developed achieved accuracy criteria in excess of the existing manual forecasting approach, while avoiding the lag time.
  • The segmentation provided deeper insights into marketing performance and instruments for optimizations.
  • The model demonstrated stability and effectiveness across different segments, with minimal performance deviations between one another.
  • Building this model helped us identify areas for the client’s product improvements, such as addressing subscription inconsistencies, expanding the dataset, and opportunities for A/B testing, which could result in even greater business value.
  • We reached all the technical goals set for the model, providing the anticipated business impact and project ROI.
Conclusion
The custom machine learning model delivered significant business value by enabling instant LTV forecasting, which improved ROAS and optimized marketing spend. Precise user segmentation and the ability to predict LTV for underrepresented segments enhanced targeting efficiency, while the model's scalability ensured long-term adaptability. The project met all technical objectives, reduced advertising waste by over 10% within the first few months, and provided a full ROI within three months after deployment, paving the way for continued growth and efficiency.