Recently, customer lifetime value (CLV) has become an accepted metric for measuring the relative attractiveness of a customer or a customer segment. Measurement issues of CLV have been addressed in a fairly large number of empirical studies, while prediction of CLV has received much less attention. In particular, the issue of predicting CLV at very early stages of the customer relationship has not been addressed. This is a topic of particular importance, since firms would benefit from a model that helps them identify customers with a high CLV even before those have proven to be high value customers. In this paper, a model is developed which aims at predicting number of transactions per period, upgrading behavior and subsequently, CLV, based on information that is realistically available to a firm at an early time of the customer relationship (customer communication, channel choice, competitiveness of choice, exhibited transaction behavior). The model is applied and tested using data from a major European airline. Results show that a) all chosen covariates help predict future behavior, b) at least to some degree, CLV predictions are possible at an early customer relationship stage and c) regular updating of CLV estimates helps improve prediction accuracy. By using information from additional data sources (share of wallet and survey data), it is also shown that customers for which CLV is overpredicted conduct a smaller proportion of their business with the firm and are less satisfied than customers for which CLV was correctly or underpredicted. Hence, model forecasts can be used to determine customers that are worth focusing on.