Challenge

Which customers are the most valuable in the long term? How can these be identified and then ideally taken into account in marketing?

Answering those questions is of particular interest to companies in retail, e-commerce, travel, insurance and subscription related businesses. More precisely, for all businesses with a short and medium purchase cycle (with max. 6 months between transactions) and a high proportion of returning customers.

Customer lifetime value (LTV or CLV) plays a central role in identifying the most valuable customers. For this concept, Google has developed the 'Crystalvalue' approach, which is implemented with the help of Vertex AI. What this approach is based on, how it works and which activation options exist are explained below.

Approach

Customer lifetime value is a metric that describes the aggregate value of a customer - ideally across the entire customer relationship. Since this value is known for history but not for the future, models are used. With these so-called customer lifetime value prediction models, the future value of each customer can be predicted. Figure 1 illustrates the basic principle of an LTV model: Based on aggregated historical 1st party data (features) during a past period (lookback window), the value (e.g. revenue) for each customer during a future period (lookahead window) is predicted. Depending on the business model and data availability, the lookback and lookahead windows are defined individually.

In addition to defining the ideal lookback and lookahead windows, the relevant historical features and the target metric are also determined. Basically, only three pieces of information per customer are required as model input: (1) a customer ID, (2) the transaction date and (3) the transaction value. This information is usually not only available in one data source; possible sources are e.g. CRM systems or Google Analytics. It is important that the available data is as complete as possible and of good and consistent quality over the relevant period. If desired, this basis of core features can be expanded to include additional features (e.g. number of items purchased, product categories, region). In this way, the model performance can finally be improved. With regard to the number of historical data, there is no fixed rule. In general, the more data, the better. For example, in order to calculate an LTV prediction for the next 365 days (lookahead window) based on the last 365 days, 2 years of data need to be provided.

Various approaches can be used for modelling. Google's Crystalvalue approach (see Figure 2) is described below. After data cleaning and preparation, an LTV model is trained and evaluated with the help of Vertex AI Auto ML. The resulting LTV model can be regularly updated via Vertex pipelines. Finally, to obtain lifetime value predictions for existing customers, the model is applied to current historical data. The resulting predicted lifetime values per customer can then be stored in a BigQuery table and made available for activation in advertising tools (e.g. Search Ads, DV360).

Result

Once the predicted customer Lifetime values are calculated, their distribution can provide valuable insights for future marketing strategy. Figure 3 shows an example distribution, where the X-axis reflects the lifetime value and the Y-axis the number of customers. Often, the Pareto principle applies in this context as well, according to which 20% of the customers (the customers with high lifetime value) are responsible for 80% of the revenue.

For further activation via marketing strategies, it is helpful to bundle the customers into different lifetime value clusters and - after a possible deeper analysis - to derive different activation strategies in the relevant marketing tools. In general, the strategies derived from the LTV approach focus on a long-term customer relationship instead of short-term profit. Possible scenarios for campaign optimisation would be, for example, improved targeting of the most valuable LTV customers, shifting the budget towards products that are particularly popular with customers with a high LTV or excluding customers with a low LTV in certain campaigns.