Convergence of outcomes and evolution of strategic behavior in double auctions

Abstract

We study the emergence of strategic behavior in double auctions with an equal number of buyers and sellers, under the distinct assumptions that orders are cleared simultaneously or asynchronously. The evolution of strategic behavior is modeled as a learning process driven by a genetic algorithm. We find that, as the size of the market grows, allocative inefficiency tends to zero and performance converges to the competitive outcome, regardless of the order-clearing rule. The main result concerns the evolution of strategic behavior as the size of the market gets larger. Under simultaneous order-clearing, only marginal traders learn to be price takers and make offers equal to their valuations/costs. Under asynchronous order-clearing, all intramarginal traders learn to be price makers and make offers equal to the competitive equilibrium price. The nature of the order-clearing rule affects in a fundamental way what kind of strategic behavior we should expect to emerge

Time-Dependent Trading Strategies in a Continuous Double Auction

“…Hence, 20 agents are randomly selected and act in a session according to a random queue. As discussed before, a relatively low N must be used to avoid trivial dynamics and we used the intermediate size of the market described in [Fano et al, 2011].…”

“…Regarding which deviations to analyze, we consider three alternative ways to sample candidates, leading to three different families of ε. First, as in [Fano et al, 2011], we examine simple constant strategies where the bid does not vary with the position in the queue. 2 Second, we create a deviation by adding a random quantity to each of the components of an existing strategy w ∈ R 20 .…”

Learning to Trade in an Unbalanced Market

“…We evolve traders' strategies using genetic programming, henceforth nicknamed GP for brevity. This optimization method is an alternative approach to the genetic algorithm used in Fano et al [2]. The design of our GP routines follows closely the standard tree-based approach described in Koza [4].…”

“…Evolution takes place sequentially over at least 2500 optimization steps. 2 In each step, one trader is randomly selected from the pool and his trading function is optimized by GP. As the complexity of the optimization problem is rather low, we choose conservative parameters for the GP algorithm.…”

Evolution and market behavior in economics and finance: introduction to the special issue

“…Fano et al (2013) investigate the emergence of strategic behavior, modeled as a learning process driven by the genetic algorithm, in double auctions comparing cases in which orders are cleared simultaneously or asynchronously. The first main finding is that, provided that the number of agents is large, the competitive outcome arises under both market architectures.…”