Customer segmentation frequently begins by identifying your most valuable customers. Most marketers – especially in DTC — start with RFM analysis. RFM is simple, intuitive, and widely used. But RFM tells you who your best customers are today, not who they will become.

That limitation becomes more pronounced as your brand scales. RFM approximates current customer value, but it doesn’t tell you where to invest next. To truly optimize retention and marketing spend, you need to move from segmentation to prediction.

Where RFM Falls Short

RFM stands for Recency, Frequency, and Monetary value. Marketers use it to segment customers based on how recently they purchased, how often they purchase, and how much they spend.

It’s highly effective for prioritization—but it’s still a snapshot.

RFM doesn’t account for how customers evolve over time, and it treats customers within the same segment as broadly similar. In reality, customer behavior is dynamic and often nonlinear. The timing of purchases, channel interactions, and engagement patterns all influence future value—but RFM doesn’t capture that.

This becomes a problem when you need to decide where to allocate your next marketing dollar.

Predicting Customer Value with Machine Learning

RFM groups customers based on past behavior. Machine learning helps you predict what they’ll do next.

Instead of asking, “Who are my best customers today?”, predictive models answer:
“Which customers are most likely to become my best customers?”

One of the most effective approaches for this is gradient boosting.

Gradient boosting builds a sequence of models – typically decision trees – where each step improves the overall prediction. The result is a model that performs well on messy, real-world data.

That matters in marketing. Customer behavior is influenced by multiple factors and often exhibits nonlinear patterns — like diminishing returns or delayed responses. Gradient boosting captures these patterns automatically, without requiring you to define them in advance.

It doesn’t assume how customers behave – it learns it from the data.

Why CatBoost Works for DTC Data

In many cases, DTC brands aren’t trying to predict an exact LTV value. Instead, they want to classify customers—for example:

• Likely to subscribe vs. not
• High vs. medium vs. low value

Historically, marketers relied on logistic regression for these types of problems. But that approach often requires extensive preprocessing – cleaning data, encoding categories, and managing missing values – which can introduce errors and slow down analysis.

CatBoost, a gradient boosting algorithm, is designed to handle categorical data natively. It reduces preprocessing effort and minimizes encoding issues, making it especially practical for DTC datasets, which are often rich in categorical variables.

The result: faster model development, more reliable insights, and outputs that better reflect real-world data.

When the Model Doesn’t Work (and Why That Matters)

About a year ago, one of CompassDTC’s clients wanted to understand the factors that drove subscriptions so they could predict which customers were most likely to convert.

The dataset was rich: hundreds of demographic, psychographic, and financial attributes across a five-figure customer base, along with purchase behavior.

We chose CatBoost given the categorical nature of the outcome and the breadth of features.

But we ran into two issues.

First, when we included transactional variables—such as recent purchase activity or customer tenure – the model performance looked unrealistically strong. That’s because these variables were already proxies for customer value. In effect, the model was “predicting” subscriptions using signals that reflected the outcome itself.

Second, after removing those behavioral variables, only demographic and lifestyle features remained. Model performance dropped significantly. In other words, demographics alone did not meaningfully distinguish subscribers from non-subscribers.

This reinforced an important reality: customer behavior – not static attributes – is the primary driver of value.

Even so, the model was still useful. It provided directional insights, relative customer rankings, and a clearer understanding of which signals actually mattered.

In real-world data, the goal isn’t perfect prediction – it’s better decisions.

Turning Predictions into Profit

As RFM – and experience – shows, not all customers deserve equal investment. Predictive models help you act on that insight with greater precision.

By applying gradient boosting techniques to predict customer value, your brand can:

• Identify high-upside customers before they fully emerge
• Allocate retention spend more efficiently
• Personalize offers based on predicted value

RFM shows you what your customers look like today. Predictive modeling shows you where they’re going.

The advantage isn’t the model—it’s how you act on it.

RFM Isn’t Dead – It’s Just Not Enough

RFM remains a valuable starting point for segmentation. Its simplicity and clarity make it an essential tool for understanding your customer base.

But it’s no longer sufficient on its own.

To compete effectively, brands need to layer in predictive models that identify future value—not just past behavior.

RFM tells you where you’ve been; machine learning helps you decide where to go next.