Abstract
This paper considers a statistical signal processing problem involving agent based models of financial markets which at a micro-level are driven by socially aware and risk- averse trading agents. These agents trade (buy or sell) stocks by exploiting information about the decisions of previous agents (social learning) via an order book in addition to a private (noisy) signal they receive on the value of the stock. We are interested in the following: (1) Modelling the dynamics of these risk averse agents, (2) Sequential detection of a market shock based on the behaviour of these agents. Structural results which characterize social learning under a risk measure, CVaR (Conditional Value-at-risk), are presented and formulation of the Bayesian change point detection problem is provided. The structural results exhibit two interesting properties: (i) Risk averse agents herd more often than risk neutral agents (ii) The stopping set in the sequential detection problem is non-convex. The framework is validated on data from the Yahoo! Tech Buzz game dataset.
Introduction
In financial markets, "shocks"—sudden jumps in asset value—are critical to detect, but they are often obscured by the behavior of the traders themselves. Traditional signal processing assumes sensors or agents are independent and static. However, financial markets are driven by social learning: traders watch the "order book" to see what others are doing, often valuing the crowd's wisdom over their own private, noisy information.
This paper develops a framework for detecting these shocks by modeling the behavior of risk-averse traders who learn from one another. By incorporating Conditional Value-at-Risk (CVaR) into an agent-based model, the authors show that risk-aversion causes traders to "herd" more quickly, which creates a complex, non-convex "stopping set" for detecting market changes—meaning the optimal time to react to a shock isn't always a simple single-threshold decision.
Methodology
The authors combine micro-level agent modeling with macro-level statistical detection through several key components:
- Risk-Averse Agent Model: Instead of standard expectation, agents minimize their CVaR, a coherent risk measure that focuses on the average loss in the worst-case scenarios.
- Social Learning Filter: The authors derive a Bayesian filter that updates the "public belief" of an asset's value based on the sequence of buy/sell actions observed in an order book.
- Quickest Detection (POMDP): The problem of detecting a shock is framed as a Partially Observed Markov Decision Process. A "Market Observer" must balance the cost of a false alarm against the cost of a delay in detecting a real shock.
- Structural Analysis: The paper proves that risk-averse agents' decisions remain "monotone and ordinal," meaning they follow intuitive rules even when they are being cautious.
Key Insights: Herding and Non-Convexity
The most significant theoretical finding is that risk-averse agents herd sooner than risk-neutral ones. Because they are afraid of the "tail" of the loss distribution, they find it safer to follow the market consensus earlier, even if their private data suggests otherwise. As risk-aversion ($\alpha$) increases, the "social learning region"—the range of beliefs where agents actually use their private signals—shrinks.
Furthermore, in standard detection, one typically stops when a belief hits a certain threshold. Here, the "stopping set" is non-convex. An observer might decide to stop (detect a shock), but if the belief shifts further, they might actually find it optimal to continue observing before stopping again. This "changing of the mind" is a direct result of how social learning masks the true state of the market.
Empirical Validation: Yahoo! Tech Buzz
The framework was validated using the Yahoo! Tech Buzz Game dataset, a real-world stock market simulation where players traded stocks based on "buzz" (search volume) for technologies like Skype and iPod. The "buzz" scores served as noisy private signals, and the transaction history provided the sequence of decisions used to test the shock detection algorithm.
Using this data, the model successfully identified a value shock in Skype stock. For a risk-aversion factor of $\alpha = 0.45$, the system detected a major shift on June 7th, just one day after the actual underlying value change occurred on June 6th.
The model achieved reasonable quickest detection performance specifically when agents were modeled as risk-averse, suggesting that standard risk-neutral models may be poorly suited for real-world financial signal processing.
Conclusion and Future Work
While the model provides a robust framework for shock detection, it currently assumes homogeneous risk-aversion across all agents. Real markets are composed of diverse participants with varying risk tolerances. Future research could explore heterogeneous risk levels and time-varying costs for market actions. Additionally, the structural results showing non-convex stopping sets provide a new perspective on Bayesian quickest detection in social learning environments.
Summary updated