-
Notifications
You must be signed in to change notification settings - Fork 7
CS7545_Sp24_Lecture_24: Optimal betting strategies, gambling and online portfolio optimization
Suppose that you are horse betting across
As it turns out, the optimal betting strategy, assuming that you have perfect information about the probability that each horse will win, is to place bets proportional to the probability that each horse will win, regardless of how large the reward is. The whole proof was not covered in class.
This does not necessarily mean that this is the best strategy. There is always the possibility of setting aside some money to not bet. As it turns out, it is possible to still lose money over time, even using this betting strategy, depending on the rewards. Alas, there is no guarantee that you will always make money, even with perfect information.
Suppose that there are
This means that on the
The Constant Rebalanced Portfolio (CRP) says that based on a distribution
Cover's Universal Portfolio (CUP) proposes to combine multiple CRPs. This method achieves polynomially worse performance than the best CRP, which is exponential. This means that CUP is only polynomially worse than the best CRP. The difference between CUP and the best CRP worsens as the number of CRPs used increases.
Proof:
Define
or in other words, the portfolio's wealth ratio.
We can treat this like an online convex optimization problem by noting that
Theorem: There exists an algorithm s.t. for any
Algorithm:
After
Lemma 1:
Let
Lemma 2:
Thus, the wealth of this strategy is at least
Letting