Dema’s profit-based lifetime value (LTV) forecasting engine transforms your historical customer behavior and transaction data into precise, forward-looking customer lifetime value predictions. These intelligent insights empower you to optimize customer acquisition costs, implement strategic retention campaigns, and allocate marketing resources with confidence—identifying high-value customers and preventing churn before they impact your bottom line.

Trained on customer specific data, the system will forecast future lifetime value (LTV) values of each customer based on the Gross Profit 2 metric.

​
How to access

You can access the LTV forecast directly from the platform’s Reports section. It is available as a selectable metric and can be explored at multiple levels of granularity—including down to the individual customer level. When applying the LTV metric to broader dimensions such as Country or Store, the platform displays the average forecasted LTV of customers who made a purchase within the selected time period. This allows you to assess expected profitability across different regions or store locations.

Because LTV forecasts are updated daily, the values shown are always the most recent predictions available, even when viewing historical timeframes. Consequently, past dates reflect the latest forecasted LTVs rather than the original values calculated at that time. This forward-looking approach ensures the data reflects your customers’ most current expected behavior.

​
How our forecasting engine works

Core algorithm

The system is built on gradient boosted decision trees, an advanced algorithm architected for large-scale customer datasets, delivering speed and accuracy for lifetime value forecasts.

Continuous model learning

The system incorporates new customer patterns, seasonal behaviors, and market changes to maintain forecasting precision throughout your business cycle with monthly training and daily prediction updates.

​
Core algorithm

The engine employs an ensemble of gradient boosted decision trees that systematically analyze your customer data through intelligent questioning patterns. For example: “Did this customer demonstrate strong engagement during recent periods?” or “How did similar customers in this demographic perform during comparable periods?”

By aggregating insights from thousands of these decision trees, the model generates comprehensive customer lifetime value forecasts. This approach provides more stable and actionable forecasts for strategic customer relationship planning.

Customer preferences and market conditions evolve rapidly. The system automatically retrains models monthly and updates predictions daily to maintain optimal forecasting accuracy as new customer behavior data becomes available.

​
Features

The system uses a combination of historical customer behavior indicators, transaction analysis, and temporal context variables to create a comprehensive feature set to forecast future profit-based LTV. The target variable employed is the Gross Profit 2 metric.

  • Customer engagement velocity and seasonal patterns
  • Geographic and demographic variations
  • Purchase frequency and monetary metrics
  • Product category preferences across customer segments
  • Channel performance and acquisition source history
  • Marketing response activity across available customer hierarchies

Customer feature importance varies significantly across different storefronts, channels and market segments. Our machine learning algorithm automatically identifies and weighs the most predictive factors specific to your unique business model and customer base.

​
Training time scales

The temporal relationship between features, labels, and forecasts is essential for constructing a robust model. The diagram below illustrates the time-based strategy used to define training and forecasting windows.

​
Training data construction

To train the machine learning model, the historical data is segmented into two windows:

Feature window (−2y to −1y): During this period, data is collected from customer behavior signals and profile attributes such as historical purchases, engagement frequency, product preferences, demographics, and other relevant metrics. These form the feature set used for training.

Label window (−1y to Present): Past the feature period, the actual LTV over the subsequent year is measured. These values serve as the ground truth labels for the model training.

This historical snapshot ensures that the model learns from the same temporal relationships it will encounter in real-world inference.

​
Live system prediction

In a live setting, the model observes the most recent data for each customer using the same feature engineering process used during training.

Prediction window (Present to +1y): The model forecasts the LTV for the upcoming year based on the most recent data. These predictions are logged for evaluation once the actual values become available. This setup ensures no data leakage and maintains realistic deployment constraints, where only past and present data are available at prediction time.

​
Technical requirements

Data requirements The LTV forecasting model requires at least 2 years of historical customer transaction data to generate reliable forecasts. This is to ensure that the model has enough data to learn the temporal patterns and relationships between features and labels.