Saturday, January 10, 2009

Idea #16 - Predicting trading pattern through data mining

Predicting the trading decision of an individual is extremely difficult, however predicting the decision of the entire universe of investor community might be slightly easier if we consider a model that takes into account of herd behavior, or information cascade, accounted for in idea #12.

Assume we have the buy/sell trading data of all individual investors and institutions for the past 20 years (well, it might be difficult to obtain, but let us assume we have access). We can very well study the decision making process of the investment society as a group, testing the accuracy or predicting power, of different mathematical models for herd behavior, under both normal trading environments, as well as extreme conditions such as in 1987, 2001 and 2008.

Probability theory tells us that the aggregate of a group of random variables with i.i.d (independent identical distribution) behaves like Gaussian, i.e. the likelihood of all individuals making the same decisions decreases exponentially. However, in a social environment, this is far from reality, i.e. an individual's decision can be heavily influenced by peers, and the likelihood of making an emulating or similar decision increases greatly as more people start to join the herd.

In other words, the probability of an event that 99% of a bond holder decide to dump it at the same time is astronomically small under Gaussian. However, it is a different story under an information cascade model. By how much? It will have to be tested by the trading history of the herd.


No comments:

Post a Comment