A No-Regret Framework For Deriving Optimal Strategies With Emphasis On Trading In Electronic Markets

Abstract:

We present a no-regret learning framework to analyze behavior of strategic agents
within single and multi-player games. We introduce a tool that could be used for
any game to calculate optimal strategies for each player. Specically, we base our
tool on two regret minimization algorithms - Multiplicative Weights [5] and EXP3 [6].

We begin by describing each of the two regret minimization algorithms used
followed by justication of our system design. Then, for proof of concept, we test
our tool on two known games. Then, given that the main intention of this work is to
measure the viability of applying a regret minimization framework within a trading
environment, we move our discussion to describing our market framework followed by testing on an abstract trading game. That trading game is one that is designed based on certain abstractions, which will be described in detail. We then evaluate the application of our framework for that trading game with future potential of applying to further trading environments.

A significant amount of effort was dedicated towards understanding the market
microstructure of an electronic market order book system. We will describe in
detail our implementation of a zero-intelligence [3] market environment, and the
implications and results for using the same. The trading game developed and tested is inherently linked to this market environment and the randomness of the market was instrumental in designing the game used.