We study the problem of testing an expert'’s forecast. We show that it is intimately related to the axiomatization of probability. Every test that admits low Type-I error can be manipulated by a strategic expert, who can pass the test with arbitrarly high probability independently of the realized outcome. If the test excludes forecasts that violate countable additivity, then there exist non-manipulable tests. We study a class of …finitely additive probability measures satisfying a natural inductive property, and we conclude that it is a sufficient assumption for restricting the effect of manipulation.