Foundations

These are some of the known facts about game-theoretic upper probability $P$:

  • It is an outer measure [obvious].
  • It is a Choquet capacity, at least in the case of a finite outcome space and one-step ahead forecasts [Vovk 2009].
  • In general, it is not strongly additive, i.e., it is not guaranteed to satisfy $P(A\cup B) + P(A\cap B) \le P(A) + P(B)$. (Therefore, the situation is similar to that in the theory of imprecise probabilities: cf. Walley (2000), page 128.) This is a simple example in the prequential framework (sequential probability forecasting of binary outcomes) with horizon 2 (i.e., the forecaster issues 2 forecasts $p_1,p_2$ for 2 consecutive outcomes $y_1,y_2$): $A = \{(0,0,1\slash{}2,0),(1\slash{}2,0,0,0)\}$ and $B = \{(0,0,1\slash{}2,0),(1\slash{}2,1,0,0)\}$ (the elements of $A$ and $B$ are represented in the form $(p_1,y_1,p_2,y_2)$). In this case we have $P(A\cup B)=1$ and $P(A\cap B) = P(A) = P(B) = 1\slash{}2$. (For the standard prequential framework with infinite horizon just add $00\ldots$ at the end of each element of $A$ and $B$.)

Bibliography

  • Vladimir Vovk. Prequential probability: game-theoretic = measure-theoretic. The Game-Theoretic Probability and Finance Project,Working Paper 27, January 2009.
  • Peter Walley. Towards a unified theory of imprecise probability. International Journal of Approximate Reasoning 24 (2000) 125 - 148.