Genetic Programming for Feature Extraction in Financial Forecasting
by József Hornyák and László Monostori
Artificial neural networks (ANNs) received great attention in the past few years because they were able to solve several difficult problems with complex, irrelevant, noisy or partial information, and problems which were hardly manageable in other ways. The usual inputs of ANNs are the time-series themselves or their simple descendants, such as differences, moving averages or standard deviations. The applicability of genetic programming for feature extraction is investigated at the SZTAKI, as part of a PhD work.
During the training phase ANNs try to learn associations between the inputs and the expected outputs. Although back propagation (BP) ANNs are appropriate for non-linear mapping, they cannot easily realise certain mathematical relationships. On the one hand, appropriate feature extraction techniques can simplify the mapping task, on the other hand, they can enhance the speed and effectiveness of learning. On the base of previous experience, the user usually defines a large number of features, and automatic feature selection methods (eg based on statistical measures) are applied to reduce the feature size. A different technique for feature creation is the genetic programming (GP) approach. Genetic programming provides a way to search the space of all possible functions composed of certain terminals and primitive functions to find a function that satisfies the initial conditions.
The measurement of goodness of individual features or feature sets plays a significant role in all kinds of feature extraction techniques. Methods can be distinguished, whether the learning/ classification/estimation phases are incorporated in the feature extraction method (filter and wrapper approaches).
In fact, most of the financial technical indicators (Average True Range, Chaikin Oscillator, Demand Index, Directional Movement Index, Relative Strength Index etc.) are features of time-series in a certain sense. Feature extraction can lead to similar indicators. An interesting question is, however, whether such an approach can create new, better indicators.
The techniques were demonstrated and compared on the problem of predicting the direction of changes in the next weeks average of daily closes for S&P 500 Index. The fundamental data were the daily S&P 500 High, Low and Close Indices, Dow Jones Industrial Average, Dow Jones Transportation Average, Dow Jones 20 Bond Average, Dow Jones Utility Average and NYSE Total Volume from 1993 to 1996.
Three ANN-based forecasting models have been compared. The first one used ANNs trained by historical data and their simple descendants. The second one was trained by historical data and technical indicators, while the third model used new features extracted by GP as well. Plain ANN models did not provide the necessary generalization power. The examined financial indicators showed interclass distance measure (ICDM) values better than those of raw data and enhanced the performance of ANN-based forecasting. By using GP much better inputs for ANNs could be created improving their learning and generalization abilities.
Nevertheless, further work on forecasting models is planned, for example:
- extension of functions and terminals for GP
- direct application of GP for the extraction of investment decisions
- committee forecasts where some different forecasting systems work for the same problem and these forecasts are merged.
This project is partially supported by the Scientific Research Fund OTKA, Hungary, Grant No. T023650.
László Monostori - SZTAKI
Tel: +36 1 466 5644