Journal on Policy & Complex Systems Vol. 2, Issue 2, Fall 2015 | Page 55

Journal on Policy and Complex Systems
from its ability to ( a ) evaluate many different rules and parameters at the individual level , ( b ) test different trust thresholds among agents , and ( c ) adjust learning aggressiveness , thus enabling the system designer and the end users to uncover agent interactions that actually improve portfolio performance .
3 . Complex Adaptive System in Investment Management

Since CAS simulations have the capability to capture the essence of distribution , self-organization , and nonlinear social and natural phenomena , characterized by feedback loops and emergent properties , they offer an innovative way of modeling inherently complex systems such as a stock market . Interaction patterns and emergent regularities are important features in the financial markets ( Cappiello , Engle , & Sheppard , 2006 ). It is natural to utilize ABM techniques to model financial markets as a dynamic system of interacting agents . There already have been successful implementations of ABM models in many theaters of human endeavor , including economics , government , military , sociology , healthcare , architecture , city planning , policy , and biology ( Hadzikadic , Carmichael , & Curtin , 2010 ; Johnson , Hadzikadic , & Whitmeyer , 2013 ). In financial market simulations , a large number of agents engage repeatedly in local interactions , giving rise to global markets ( Bonabeau , 2002 ; Darley & Outkin , 2007 ; Roberto , Cincotti , Focardi , & Marchesi , 2001 ). The dynamics can be readily captured by a well-designed and CAS-enabled ABM simulation . ABM models can be augmented with learning , as demonstrated by Farmer ’ s agent-based investment model ( Farmer , 2001 ) that yields powerful insights about market behaviors . Also , ABM models combined with artificial intelligence-based reinforcement learning provide a plausible way of stock modeling ( Ramanauskas , 2008 ).

In this paper , we describe an ABM system that seeks a balanced level of trust and learning aggressiveness for individual agents , based on a single stock-trading model presented by Su and Hadzikadic ( 2015 ). This model issues a stock-trading signal ( buy , sell , or hold ) for a stock ( Bank of America , BAC , in our example ) on a daily basis . Agents trade the BAC stock based on the publicly available , adjusted , daily data from January 2 , 1987 to December 31 , 2014 .
4 . A CAS Stock-Trading Model

We built a stock-trading model using the concept of CAS that issues a

stock-trading signal at the end of each trading day . Agents use the current closing price of BAC to determine the next trading action , which includes buy , sell , and hold options . The closing price is adjusted to eliminate the effect of stock dividends . Interest is distributed based on the agent ’ s cash on hand
52