In the dynamic landscape of mobile gaming, e-commerce, and subscription businesses, leveraging predictive models to optimize user acquisition (UA) efforts is crucial for achieving sustainable growth and maximizing long-term return on ad spend (ROAS). The two main UA-related tasks are:

1. Ensuring a balanced mix of paid traffic sources.

2. Ensuring that ad platforms attract users with high LTV.

Many companies face challenges when solving both tasks using the same predictive model. In this article, we argue why two separate models are required. We outline the key differences between cohort-based LTV models (type 1) and user-level pLTV models (type 2), their distinct applications, and why using both would better serve these UA-related tasks.

Cohort-based LTV models predict the collective LTV of groups of users (cohorts) acquired during a specific period. These models aggregate data from multiple users to provide an average LTV prediction for the entire cohort. They typically use a collection of data timeframes and horizons as input and output, with common pairs being D0->D7, D7->D30, and D7->D180. This model type is more common, and many businesses have developed or plan to develop it internally.

The relative ease of implementing cohort-based LTV models internally is due to the availability of off-the-shelf products, platforms, and a wealth of knowledge in the market. A company may start with a model as simple as curve fitting or RFM and get reasonable results. Even for more advanced ideas, external resources are abundant. Projects like Google Cloud's BigQuery ML, internal resources in MMPs such as Singular, and various GitHub projects make it accessible for businesses to build and deploy these models effectively.

Some of Churney’s customers have invested considerable internal resources to develop these models. As data scientists, we recognize and appreciate the value of this work, and you can still benefit from working with us even if you have developed these models in-house.

Budget Allocation: Cohort-based models are useful for evaluating the performance of ad campaigns and making decisions about budget allocation at a higher level.

Performance Reporting: These models help businesses understand the overall effectiveness of their marketing strategies by providing insights into the long-term value of user groups.

User-level pLTV (predicted Lifetime Value) models focus on predicting individual users' future value. By analyzing historical user data and behaviors, these models forecast how much revenue a specific user will likely generate over their lifetime or by a certain future day. This granular approach allows businesses to make precise predictions at the user level, which can be incredibly powerful for optimizing UA campaigns.

Simple user-level models also benefit from open-source community support. However, these models may sacrifice performance if they consider only some available data. Feature-based models rely heavily on specific attributes or features of a product to make predictions. When the product undergoes changes, these models can become less accurate because the features they depend on may no longer be relevant or consistent. This sensitivity to changes can lead to inaccurate predictions and reduced performance. Deep learning models, which analyze all data, can overcome this issue but are more complex to implement.

Personalized Marketing: User-level insights enable more personalized marketing efforts, increasing user engagement and retention.

Targeted Bidding: By predicting the future value of individual users, businesses can bid selectively on high-value users, optimizing their ad spend.

It's a common misconception that a user-level pLTV model can replace a cohort-based model for all use cases or vice versa. However, these models serve different purposes and are not interchangeable due to the following reasons:

User-level predictions by type 2 models are highly specific and may introduce inaccuracies when aggregated to the campaign or other levels. Cohort-based models provide more accurate insights for budget allocation and performance evaluation. Aggregating user-level predictions can lead to misleading results, as individual variances may distort the overall picture.

Cohort-based models often rely on data available seven or more days after a conversion event, which may need to be more timely for close to real-time UA optimization. In contrast, user-level pLTV models can provide much faster predictions, allowing for more agile and responsive UA strategies. This timeliness is critical, especially given the ad platform's decay in processing and optimizing based on incoming events after approximately 24 hours from the acquisition event.

Implementing user-level pLTV models for UA targeting optimization involves complex activation processes, including deciding which signals to send to the ad platform, timing these signals, and navigating the specific intricacies of different networks. This requires a deep understanding of UA optimization's technical and strategic aspects.

In contrast to cohort-based models, user-level pLTV models should be based on causal deep-learning models to avoid drift and ensure optimal performance at scale over time. The causal element allows businesses to make changes while the model delivers accurate predictions, remaining agnostic to any product, game, or offer modifications.

Both user-level pLTV models and cohort-based LTV models play critical roles in optimizing UA efforts. Understanding their differences, applications, and limitations is key to leveraging their full potential. At Churney, we have deep causal ML pLTV models tailored specifically for UA optimization. These models bring in the best users possible and increase ROAS at scale. Adopting the right model for the right task can significantly improve your marketing efficiency and drive better results for your business.