The Solomonoff-Levin theory of universal prediction relies on a Bayes-type merging procedure that is applied to a large class of prediction strategies (such as all computable strategies). The desirable properties that we would like universal predictions to satisfy can be roughly classified into theoretical (as many as possible of the standard laws of probability should hold for the predictions and the actual observations) and pragmatic (the decisions based on the predictions should lead to a small loss and/or large gain). Several such properties are known to be satisfied by the Solomonoff-Levin predictions, but it turns out that a less direct approach to universal prediction leads to many more properties of both kinds. A new player is added to the game of prediction: the gambler, who tries to make money on the lack of agreement between the predictions and the actual observations. A large class of potential strategies for the gambler can now be merged into one "universal law of probability", and it can be shown that there is a way of producing predictions that are ideal as far as this universal law is concerned. Such predictions can also be regarded as universal. They can be shown to satisfy the standard properties of calibration and resolution and to give good predictions in the framework of competitive on-line learning.