We intoduce a new origin of volatility clustering in econonmic time series gererated by systems of interacting adaptive agents. Each agent is assigned a random subset of a fixed collection of predictors. At every time step each agent generates an action based upon its assigned predictors. Some agents are contrarians---i.e. they act at variance with the natural action suggested by a predictor. Agents that perform poorly are replaced. At each time step the signal value is generated soley by the cumulative actions of the agents on the current history of the time series. We observe numerically that under the dynamics induced by the removal of poor performers, even when contrarians are introduced at a very low density, the system evolves to a state in which contrarians comprise nearly half of the population. Furthermore, the time series generated by these systems exhibits volatility clustering. Elimination of either the contrarian behavior or the removal of poor players precludes volatility clustering.
[1]
R. Engle.
Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation
,
1982
.
[2]
W. Arthur.
On Learning and Adaptation in the Economy
,
1992
.
[3]
Bernardo A. Huberman,et al.
Clustered volatility in multiagent dynamics
,
1995,
adap-org/9502006.
[4]
R. Chou,et al.
ARCH modeling in finance: A review of the theory and empirical evidence
,
1992
.
[5]
V. Akgiray.
Conditional Heteroscedasticity in Time Series of Stock Returns: Evidence and Forecasts
,
1989
.
[6]
G. Schwert.
Why Does Stock Market Volatility Change Over Time?
,
1988
.
[7]
W. Arthur.
Inductive Reasoning and Bounded Rationality
,
1994
.