knowledge transfer


Fuzzy Exception Learning to Detect Noise Trading

Submitted by Uzay Kaymak, 23.01.2002, IBA-E


Noise trading is known in financial literature as agents' market activities that are not rationally based on the arrival of new information about asset values. Technical traders, who forecast financial price movements based on past prices and volumes are likely candidates for noise traders. Technical traders bid up prices in a bullish market and bid down prices in a bearish market, resulting in shorter or longer periods of serial correlation in returns. Timely detection of the occurrence of these incidental periods offers extra arbitrage (trade) opportunities to traders. Hence, these exceptional situations need to be detected before they can be acted upon in trading decisions.


Periods where noise trading has significant influence on price formation are exceptional. They have unequal lengths and appear at irregular moments, often with gradual (i.e., fuzzy) regime transitions, which make them difficult to detect by conventional statistical analysis. Erasmus University Rotterdam has developed a method based on fuzzy set theory for detecting the gradual regime transitions. Gradual regime transitions can be described naturally within the framework of fuzzy set theory, amongst other by using linguistic rules that can be verified by experts. The fuzzy exception learning method developed observes the average behavior of system outputs and tracks deviations from this average behavior. These deviations are then correlated to regions within the system's input space. The result is a set of fuzzy rules that describe the regimes, which lead to deviations from the average system behavior. Temporary deviations from the average system behavior due to noise trading can thus be detected.

Status and results

A prototype system is available. The method has been applied to both real and simulated financial data, and it has been found to detect regime shifts with sufficient accuracy for simple financial products. Some of these regime shifts could be attributed to noise trading. However, the influence of the detection method on actual trading decisions is unknown at the moment. Recently, it has also been shown that the fuzzy exception learning algorithm can be explained in terms of probabilistic fuzzy systems. In the future, the link to other probability-based methods will be investigated.

Adaptivity and portability

The training algorithm for fuzzy exception learning can be both incremental and batch. The system demonstrates type-I adaptation in the former case, since the rules are modified as new data comes in. However, this is a very simple form of adaptation. Once the regime shifts are detected, the trading decisions can be made adaptive to the dominant regime at that moment. This is again typically type-I adaptation. Note that it is possible to have a mixture of regimes, some of which are coming up gradually while others fade away, which warrants the gradual transition from one regime to the other. It is also possible to model abrupt transitions between regimes. The solution is not portable at the moment. It could be a future research project to study how the model for the regime shifts for one product can be extended to predict the regime shifts for another product. There should be sufficient possibilities in this direction, since the price developments of different financial products are related to one another's through the market mechanism.

More information

W. M. van den Bergh, J. van den Berg and U. Kaymak. Detecting noise trading using fuzzy exception learning. In Proceedings of Joint 9 th IFSA World Congress and 20 th NAFIPS International Conference, pages 946-951. Vancouver, Canada, July 2001.

J. van den Berg, W. M. van den Bergh and U. Kaymak. Probabilistic and statistical fuzzy set foundations of competitive exception learning. In Proceedings of the Tenth IEEE International Conference on Fuzzy Systems, volume 3. Melbourne, Australia, December 2001.

technical reports & case studies

executive summaries

scientific papers




search eunite with