Agent-based Simulation of Financial Markets

1
Agent-based Simulation of Financial Markets

• Ilker Ersoy

2
Introduction

• Simulation of financial markets is a new fast
  growing research area.
• Main motivations are
• to provide a testbed for automation of financial
  markets
• To provide thought experiments to understand the
  moods of markets which cant be explained by
  Rational Expectations theory.

3
Markets as Complex Systems

• Financial markets are complex systems with
  behaviors such as bubbles and crashes.
• This complexity defies traditional mathematical
  analysis.
• The complexity arises from the interactions and
  expectations of the agents (buyers,sellers,etc.)
  in the market.
• Agent-based market simulation is one of the
  applications of Artificial Life.

4
Rational Expectations Theory

• Conventional RE theory assumes that
• Agents deduce their optimum behavior by logical
  processes.
• Agents have full knowledge of the market.
• Agents know that others work with same knowledge
  on the same rational basis.
• These assumptions are too strong, in most cases
  simply not true for financial markets.
• RE theory does not explain dynamic behavior of
  markets.

5
Agent-based Simulation

• In a real financial market, there are
  heterogeneous agents with different expectations
  and different levels of knowledge.
• Agent-based simulation takes this approach to
  create an artificial market.
• Agents start with little rationality and
  specialized knowledge and adapt or learn becoming
  experts in their domains.

6
Agent-based Simulation

• Advantages
• None of the assumptions of RE theory is required.
• Even the modeler does not need to have the
  knowledge to derive an optimum solution for each
  agent.
• This approach is inductive not deductive, this is
  much closer to normal human behavior.

7

• Advantages
• This approach is applicable in situations where
  RE theory produces no answers due to lack of
  single well-defined equilibrium solution.
• This approach can predict and interpret dynamical
  behaviors, not only the final outcome.

8

• Disadvantages
• Lack of analytic methods, it is largely
  computational.
• Multitude of possible algorithms for learning and
  adaptation.
• Sensitivity to parameters such as learning rate.

9
Implementation

• Agents use learning classifier systems (LCS) to
  gather knowledge and assess their rules.
• Each rule is assessed by the outcome of the
  execution of that rule (gain or loss).
• Rules are eliminated by genetic approach, new
  rules are created by mutation and crossover.

10
Implementation

• LCS classifies the environment (market) into
  classes.
• Each bit represents the existence of a certain
  condition in the market or for a stock.
• Agents place their bids for stocks, and price is
  established according to supply and demand.

11
Experiments

• A number of experiments can be conducted in this
  setting.
• Different agents can be created representing
  different investor classes for a realistic market
  simulation.
• Experiments show that even this simple simulation
  is able to produce complex dynamic behavior such
  as bubbles and crashes.

12
Problems and Future Directions

• Establishing stock prices need a realistic
  clearing mechanism, not studied broadly by far.
• LCS is suitable for genetic approach but might
  not be suitable to represent realistic knowledge
  about market.
• Calibration to real markets should consider the
  fact that investors have a long history of
  knowledge of the market to learn from.