CLV Ultra: Revolutionizing the Business Case for Customer Lifetime Value (CLV) Modeling

Customer Lifetime Value (CLV) has evolved from a buzzword that is popularly discussed but is often lacking in both substance and validity, to a critical metric that companies across industries are keen to integrate into their decision-making processes, serving as a key tool to drive customer centricity and optimize profitability. However, traditional publicly available models often fall far short in providing the accuracy required to have the conviction to actually take action – to put your company’s hard-earned capital at risk – based on CLV. Recognizing this challenge, Theta has taken the lead in addressing the shortcomings of classical CLV models head-on, creating a new model, CLV Ultra, that offers up new levels of accuracy and actionability for marketers, executives, and investors. 

Our journey began by developing a best-in-class CLV model that addressed some of the most important limitations of publicly available models. For instance, we integrated time-varying covariates, enabling us to more precisely capture and control for seasonality, market shifts, external demand shocks, customer lifecycle effects, and more. We also made significant strides forward on the “cold start problem” – the fact that traditional CLV models struggle with customers with very little or even no historical transaction data – by intelligently weighting more recently acquired customers’ own data with data from their more tenured peers. Furthermore, from a computational perspective, we fine-tuned our models to accommodate datasets with hundreds of millions of customers in an incredibly efficient manner – requiring just a few minutes to run the model on the largest datasets imaginable. And there’s much more – you can read more about how our models differ from other CLV models in this blog post

While our initial effort yielded a very high level of accuracy and diagnostic value, it required a skilled data scientist to specify and validate the model, requiring human judgment in the covariate setting process (inevitably leading to some level of performance loss), longer turnaround times and, consequently, bigger price tags. Thus, CLV Ultra was born — a breakthrough that marries unparalleled accuracy with seamless automation. Leveraging advanced machine learning techniques and predictive customer behavior models, CLV Ultra detects, decomposes, and estimates covariates with unprecedented precision, granularity, and speed. 

However, while accuracy and speed are important, their true value is contingent upon their ability to generate higher profits. On this point we want to be absolutely clear – these aforementioned advancements are much more than just nice-to-have incremental technical improvements; they “change the game” for how teams can act upon and enhance CLV. 

Benefits for managers and executives

Here are some of the managerial use cases enabled by CLV Ultra for businesses and investors.

More profitable customer acquisition. CLV Ultra fits one model on your entire customer base, enabling much better predictive accuracy for recently acquired customers by allowing them to “learn” better from more tenured customers. And while our model automatically incorporates a wide range of covariates “out of the box,” it is also much easier for the new model to proactively incorporate a broader array of covariates, not only to account for events that occur in calendar time, but also features of your customers that may be diagnostically relevant for your business – behavioral data, geographical information, demographic and psychographic variables, and the like. This greatly improves our ability to understand – and accurately estimate ROI! – for different, more fine-grained acquisition characteristics, such as channel, campaign, geography, or other funnel activities. You will get a more accurate read on how much profit you will generate by investing in one acquisition channel versus another, which will help you earn a much higher rate of return on your customer acquisition budget. If you are highly uncertain about these ROI figures, which is likely the case if you are using traditional CLV models, you cannot confidently invest in customer acquisition based on long-term value creation. 

More (and better) customer retention and development opportunities. You probably have many strategically and tactically relevant questions that have historically been unanswered that relate to the value of your existing customer base: 

  • How does customer satisfaction, or changes in it over time, relate to long-term customer value? 
  • How effective was that last mailing campaign, not in terms of short-term revenue generation but rather in terms of long-term value creation, and which type of customers had the most long-term value expansion? 
  • How does the long-term value of acquired customers located near physical store locations or distribution centers differ from customers who live further away?
  • For marketplace businesses, how is long-term consumer demand impacted by marketplace supply, and how might those impacts be different for some customers versus others? 

These and related questions are often very difficult to answer but very important – they may influence how much to invest (and to whom) in that next latest and greatest CX campaign, whether and where to place that next retail store location, whether to double down (or conversely, eliminate) the next mailing campaign, and where to drive investments on the supply-side of your marketplace. 

The reason that these questions are traditionally very difficult to answer is because traditional CLV models are physically unable to provide you with an answer. Your model would need to be accurate over long forecasting horizons, not only at the cohort level, but all the way down to the microsegment level. And importantly, your model would need to be able to discern where and in what way customer response differs based on a strategically important factor like customer satisfaction, the details of a mailing campaign, or the depth of marketplace supply for some customers versus others. Traditional CLV models can’t proactively incorporate the covariates necessary to uncover these relationships and thus try to do so retrospectively, which significantly weakens individual-level predictive accuracy. Other models are physically able to incorporate covariates proactively, but doing so requires a Herculean effort, or have other significant drawbacks (additional technical comments on this at the bottom of the post).  

Our model is uniquely suited to these use cases because it is highly accurate and as we discussed above, exceptionally accommodating of those strategically important variables. 

Easier to embed into business operations. To be able to generate the most profit from a CLV model, your CLV model needs to be embedded within the organization, supporting repeatable business use cases. It needs to be more than a source of high-level insight every once in a while – this can be helpful and a good place to start, no doubt, but you are likely leaving money on the table. 

One of the other game changing aspects of CLV Ultra is its speed and automation. The time it takes to go from data to CLV estimates is significantly lower, often by a factor of 2-3 times. This means that you won’t have to wait longer on model updates to be able to get the CLV results that power those customer acquisition, retention and development benefits that we allude to above – in fact, you’ll likely be waiting a lot shorter

We have also created an API that allows you to programmatically query for exactly the data that you need, and to get that data where you need it (e.g., into your CDP or email automation tool). You won’t have to do the full “data dump” each time you want access to the latest results for the customers targeted for that mailing campaign – you can pull just what you need for these specific customers to support that next campaign.

Between the new model and our API, you can have access to the results you need, when you need them, where you need them. 

Give it a try through our pilot program!

We hope this gives you a better sense of why we are so excited about CLV Ultra. Yes, it has many technical benefits – higher accuracy, more interpretability, faster speed, and more. But the real reason these technical benefits are important is because they enable a world of new profit creating business opportunities for managers and executives. More profitable customer acquisition, a wider “range of managerial motion” in terms of customer retention and development, and much better ease-of-use to allow you to operationalize the model and drive improvements in your day-to-day tactical decision-making.  

With that, we’re thrilled to promote our Early Access CLV Ultra Pilot Program, set to revolutionize the business case for CLV. Secure your spot today and gain access to a dashboard, detailed individual-level results, pre-defined audience segments, and tailored suggestions and use-cases for enhancing acquisition, retention, and development strategies. Don’t miss this opportunity to propel your ROI and profitability forward – contact us now to reserve your seat! To apply for the pilot program, click here. If you have any questions about the offering and whether it would be a good fit for you, please reach out to us at

Technical Comments for Data Scientists and other Geeks

The drawbacks of traditional ML approaches. You may be wondering: “CLV models without covariates? Our random forest model eats covariates for breakfast!” Yes, if you are using some sort of ML model for CLV, you may have a lot of covariates in your model, but your model may be ignoring the data from your less tenured customers (because discriminative ML models need to observe the “long-term outcome” for a customer to be included in model training, and that long-term outcome hasn’t been observed yet for recently acquired customers by construction). This leaves these models highly vulnerable in situations where customer dynamics may have changed for the less tenured customers versus the more tenured ones. 

It would be training on all customers except those that are most relevant for customer acquisition – the more recently acquired ones! 

These models also tend to do a poor job of capturing the flow of purchase activity over time, which also make it hard to rely on these models for long-term value prediction. 

Don’t get fooled by accuracy metrics on old cohorts. Predictive accuracy with more tenured customers does not mean predictive accuracy with less tenured customers. You may predict well for customers that have been around for two years, but you likely predict much worse for customers acquired within the past three months. You owe it to yourself to re-analyze your model’s performance on these customers. If performance is poor for these “young” customers, you can’t really trust your CLV model for customer acquisition use cases because you can’t trust the ROI estimates coming out of them for the customers that are most relevant for customer acquisition.