Herd Behavior in Financial Markets: Theory and Evidence

Dr. Antonio Guarino

According to anecdotal evidence, herd behavior is diffused in financial markets. Some agents buy stocks simply "following the herd" in a bull market, or sell them in the wave of diffused pessimist in a bear market. While there is a large anecdotal evidence on herd behavior in financial markets, and the idea that traders herd is diffused among practitioners and market participants, our understanding of why herding can arise in the financial market is still limited. Furthermore, studies that have attempted to document and quantify the presence of herding in the market have obtained results that are far from conclusive. Most work remains to be done both at a theoretical and at an empirical level. My research activity is devoted to shed light on the phenomenon of herd behaviour in financial markets with a combination of theoretical, experimental and empirical work. Most of this work is a joint effort with Marco Cipriani , of George Washington University.

Herd behaviour is relevant for the financial market informational efficiency. Essentially, herd behaviour entails neglecting private information in trading a stock or a group of stocks, because of what previous traders have decided to do. Clearly, if an agent neglects private information and herds, the market will not be able to infer that private information from the choice of that agent, and the price will not be able to reflect it. If many traders herd, the stock price can be misaligned with respect to the true value of the stock. Cipriani and I have proven that, indeed, even if traders are perfectly rational Bayesian agents, herd behaviour occurs in the market because of traders' heterogeneity. Traders go to the market to exploit their private information and for non-informational (e.g., liquidity) motives. Over time, as trade proceeds, the informational reasons to trade become less important than the non-informational ones. At a certain point, all traders will neglect their private information and an "informational cascade" (i.e., a situation in which the traders' decisions are completely uninformative on the asset value) will arise. In a cascade the price does not react to the order flow and is stuck for ever. It may well be that the price remains stuck at a level far from the fundamental one. For instance, it may be that the value of a stock is high, but the price remains stuck at a low level, after a sequence of sell orders in the market. Therefore, herd behaviour can be a reason why we can observe a financial crisis even in an economy with sound fundamentals.

To test theoretical models of herd behaviour is difficult. As I said, herding, at least in our framework, means neglecting private information because of the observations of previous traders' decisions. Therefore, ideally, to test such models one would need data on private information, which is of course difficult to obtain for actually running markets. This difficulty can be overcome in an experimental study. In an experiment we can observe variables not available for actual markets, in particular, the private information that agents have when making their decisions. For this reason, Cipriani and I have run a series of experiments in which we test whether herding arises as theory predicts or not. In our laboratory market, subjects receive private information on the value of a security and observe the history of past trades. Given these two pieces of information, they choose, sequentially, if they want to sell, to buy or not to trade one unit of the asset. By observing the way in which they use their private information and react to the decisions of the previous traders, we can directly detect the occurrence of herding. Our experimental results are encouraging for the theory. Although the laboratory reveals some anomalies in subjects' behaviour, i.e., some deviations from equilibrium behaviour, overall what we observe in the laboratory is quite consistent with the theoretical predictions. In particular, in a frictionless financial market in which traders trade for informational reasons only, and the price is efficiently set by a market maker in response to the order flow, herd behaviour is very limited. This agrees with the theory, according to which we should not observe herding at all. Market frictions, instead, generate informational inefficiencies both theoretically and experimentally. Cipriani and I have studied a market in which such frictions take the form a trade cost. We have shown that such a cost impairs the process of information aggregation and an informational cascade occurs almost surely. A cascade occurs since, after some trades, the gain form trading becomes lower than the cost, and agents prefer to abstain from trading. Interestingly, such behaviour is not only a theoretical result, but also occurs in the laboratory. Indeed, our experimental results are very close to the theoretical prediction. Transactions costs can be due, for instance, to transaction taxes, like the Tobin Tax. Therefore, our work has also some policy implications on the introduction of such a tax.

In more recent work, Cipriani and I have extended our research on herd behaviour by estimating a theoretical model of herding with field data. While in previous work we have brought the model to the laboratory, in this work we have brought the model to the real financial market. We use data on many days of trade for some specific stocks. We estimate the parameters of a theoretical model that predicts herd buying (or selling) when a sufficiently high sequence of traders have decided to buy (or sell) the asset. In the market there is a market maker who efficiently sets the prices for a stock according to the orders flow, and a sequence of traders, who can buy or sell at those prices. Herd buying arises when all traders value the asset more than the price posted by market maker, even if they have negative private information. Similarly, herd selling arises when all traders value the asset less than the market maker, even if they have positive private information. After we have estimated the parameters of the model, we are able to track down the beliefs that the market maker (the notional prices) and the traders have during each day of trading. By comparing these beliefs, we can identify the periods of the trading day in which herding occurs or not.

The research activity described so far refers to the analysis of a single market (for instance, a single stock, or a single country). An important development of this research is to consider many markets at the same time. This allows to studying phenomena of "financial contagion" due to learning or herding across markets. Cipriani and I have already given some theoretical contribution on this topic, but much more work needs to be done. Future research (joint with Steffen Huck) will be focused on understanding financial contagion due to informational spillovers in the laboratory. This research project is kindly supported by the ESRC (program on the World Economy and Finance).

While most of this research project on herding and learning focuses on financial markets, part of it is devoted to more general issues of "social learning." Herd behaviour in financial markets is a possible outcome of the traders' learning process. Traders try to learn from the trading decisions of the agents who went first to the market: it may happen that this learning process makes them just neglect their private information. This is what can happen in different social contexts too. When we have to decide which car to buy or which restaurant to go to, we do observe the decisions of other agents, for instance of our neighbours. After observing them, we may well decide to change our mind and, for instance, imitate their decisions, neglecting our original private information. The process of learning from the actions of others is what we call social learning. I am investigating issues of social learning not strictly related to the financial market in a number of papers. In a joint paper with Steffen Huck, and Thomas Jeitschko, we experimentally investigated a model of Bayesian learning that predicts sudden changes in the behaviour of agents and sudden crashes in the market. In the laboratory these changes (information avalanches) do not take place since decision makers are prone to a "solipsism bias," i.e., they are unable to imagine that others who play the same game may make quite different experiences than they themselves. We conjecture that this solipsism bias is of relevance in many fields of human decision making. In another work with Steffen Huck and Heike Harmgart, we analyze how agents learn by observing the number of agents who have already made a decision. For instance, agents have to decide to which of two restaurants to go, and they can infer the quality of the restaurant by some private information and by observing how many other agents have already chosen to go to the two restaurants. The question is whether the aggregation of information is efficient in this context or whether informational inefficiencies can arise. This work, still in progress, is both theoretical and experimental. If you are interested in this work, you can also look at Steffen Huck's research on social learning . In another project with Antonella Ianni we analyze the process of social learning in an economy in which agents can only observe the actions of their neighbours. The question is whether local interaction leads to complete learning in the entire society and whether convergence is fast or slow.

If what you have read so far seems interesting to you, and you want to know more, please visit the ELSE Working Papers Archive and my Personal Webpage, where you can find updated versions of my papers. adapted from http://else.econ.ucl.ac.uk/newweb/research/blurb1.php