Computational Finance 2004 Artificial Agents and Speculative Bubbles Y. Semet , S. Gelly, M. Schoenauer, M. Sebag Optimization & Machine Learning Group INRIA Futurs Orsay, France
Roadmap Agent-based Computational Economics (ACE) A financial asset trading problem: speculative bubbles A simple model studied in two contexts: Exogenous risk, ZI, finite time horizon; Endogenous risk with elementary strategy. A set of conditions for the appearance of speculative bubbles
Background European Project: DREAM (Distributed Evolutionary Algorithms…) Social and Economic problems are interesting to AI: high frequency data, distributed interactions, stochasticity, cognition and rationality, emergent computation. Market efficiency: simple models for a critical debate.
Previous Works Gode & Sunder, 93: Rationality and market efficiency Tool: « Zero-Intelligence » traders Conclusions: aggregate rationality The Santa Fe Artificial Stock Market, 96 Complexity and inductive reasonning Tool: Genetic Algorithms Conclusions: a hint on the efficiency/speculation boundary
Our market Double Auctions, Cleared Book Convention 1 financial asset 2 kinds of Agents: ZI and  Agents are randomly endowed with cash & shares
Auction Algorithm
ZI-Agents, Exogenous Risk Risk set by given external function Simulates decreasing hope for dividend collecting Finite time horizon
Evaluating Risk Endogenously Risk is high when: P is far above F P is going down Weights = control A smoothing sigmoïd r(i,t): risk at time t for agent i p(t): asset price a, v, w: controlling  weights   F: fundamental value
Choosing a strategy… 1 1 1 Buy Idle Sell 0 Exuberance Buy Idle Sell 0 Comfort Buy Idle Sell 0 Panic Exuberance Comfort Panic  R 0 R 0 0 1 r(i,t)
Making an offer Anchoring effect: offers are uniformly distributed around previous price. In most cases: An asymetric possibility for the panic mode:
A glimpse on the GUI Code in JAVA Large number of control parameters An even larger number of time series Visualization is a critical issue for ACE experimetns
Experiments Value Tuning + Heterogenization Many cross dependencies Reminder: w, alpha, R0, F, Pricing policy. Typical values:… Around 1 minute of computation time, grows very quickly with # of agents
Linear risk ; Time horizon=500 p(t) Buy & Sell r(t)
Endogenous risk: default behavior (efficiency) p(t) r(t) Buy & Sell Exuberance Comfort Panic
At t=250, F becomes 75 p(t) r(t) Buy & Sell Exuberance Comfort Panic
Bubbles w/o krach ; R0 in [0.4;0.6] p(t) Close-up Buy & Sell Exuberance Comfort Panic
Bubbles w/o krach ; R0 in [0.4;0.8] p(t) Buy & Sell Exuberance Comfort Panic
Influence of fool factor 1.05 1.1 1.2 1.0
Introducing asymetry p t-1 -5% in panic vs p t-1  +/- 1% comfort/exb.  bubbly behavior (R0 in [0.4;0.6]) p(t) Close-up Buy & Sell Exuberance Comfort Panic
Bubbly behavior (R0 in [0.4;0.8]) p(t) Close-up Buy & Sell Exuberance Comfort Panic
Conclusions Speculative bubbles in two contexts With exogenous risk Endogenous risk in conjunction with: High sensitivity to recent trends Biased heterogeneity Asymetric pricing strategy Future work: a Game Theory take Contact: semet@lri.fr

Slides Cf04

  • 1.
    Computational Finance 2004Artificial Agents and Speculative Bubbles Y. Semet , S. Gelly, M. Schoenauer, M. Sebag Optimization & Machine Learning Group INRIA Futurs Orsay, France
  • 2.
    Roadmap Agent-based ComputationalEconomics (ACE) A financial asset trading problem: speculative bubbles A simple model studied in two contexts: Exogenous risk, ZI, finite time horizon; Endogenous risk with elementary strategy. A set of conditions for the appearance of speculative bubbles
  • 3.
    Background European Project:DREAM (Distributed Evolutionary Algorithms…) Social and Economic problems are interesting to AI: high frequency data, distributed interactions, stochasticity, cognition and rationality, emergent computation. Market efficiency: simple models for a critical debate.
  • 4.
    Previous Works Gode& Sunder, 93: Rationality and market efficiency Tool: « Zero-Intelligence » traders Conclusions: aggregate rationality The Santa Fe Artificial Stock Market, 96 Complexity and inductive reasonning Tool: Genetic Algorithms Conclusions: a hint on the efficiency/speculation boundary
  • 5.
    Our market DoubleAuctions, Cleared Book Convention 1 financial asset 2 kinds of Agents: ZI and Agents are randomly endowed with cash & shares
  • 6.
  • 7.
    ZI-Agents, Exogenous RiskRisk set by given external function Simulates decreasing hope for dividend collecting Finite time horizon
  • 8.
    Evaluating Risk EndogenouslyRisk is high when: P is far above F P is going down Weights = control A smoothing sigmoïd r(i,t): risk at time t for agent i p(t): asset price a, v, w: controlling weights F: fundamental value
  • 9.
    Choosing a strategy…1 1 1 Buy Idle Sell 0 Exuberance Buy Idle Sell 0 Comfort Buy Idle Sell 0 Panic Exuberance Comfort Panic  R 0 R 0 0 1 r(i,t)
  • 10.
    Making an offerAnchoring effect: offers are uniformly distributed around previous price. In most cases: An asymetric possibility for the panic mode:
  • 11.
    A glimpse onthe GUI Code in JAVA Large number of control parameters An even larger number of time series Visualization is a critical issue for ACE experimetns
  • 12.
    Experiments Value Tuning+ Heterogenization Many cross dependencies Reminder: w, alpha, R0, F, Pricing policy. Typical values:… Around 1 minute of computation time, grows very quickly with # of agents
  • 13.
    Linear risk ;Time horizon=500 p(t) Buy & Sell r(t)
  • 14.
    Endogenous risk: defaultbehavior (efficiency) p(t) r(t) Buy & Sell Exuberance Comfort Panic
  • 15.
    At t=250, Fbecomes 75 p(t) r(t) Buy & Sell Exuberance Comfort Panic
  • 16.
    Bubbles w/o krach; R0 in [0.4;0.6] p(t) Close-up Buy & Sell Exuberance Comfort Panic
  • 17.
    Bubbles w/o krach; R0 in [0.4;0.8] p(t) Buy & Sell Exuberance Comfort Panic
  • 18.
    Influence of foolfactor 1.05 1.1 1.2 1.0
  • 19.
    Introducing asymetry pt-1 -5% in panic vs p t-1 +/- 1% comfort/exb. bubbly behavior (R0 in [0.4;0.6]) p(t) Close-up Buy & Sell Exuberance Comfort Panic
  • 20.
    Bubbly behavior (R0in [0.4;0.8]) p(t) Close-up Buy & Sell Exuberance Comfort Panic
  • 21.
    Conclusions Speculative bubblesin two contexts With exogenous risk Endogenous risk in conjunction with: High sensitivity to recent trends Biased heterogeneity Asymetric pricing strategy Future work: a Game Theory take Contact: semet@lri.fr