This report summarizes trading simulations we run on several stocks using KnowledgeMiner's prediction capabilities. Its intention was to integrate self-organizing data mining into financial decision making by applying the idea of predictive control to some known trading indicators.

Today there are the following demands on financial analysis:

• financial markets are integrated. An intermarket financial analysis needs adequate tools for modeling of complex systems, where all factors are influenced by a range of other factors: some are known, some are unknown, some are quantifiable, some are objective and some are subjective [Kindgon, 97].
• many rules that describe the underlying financial, economical processes are qualitative or fuzzy requiring judgement, and therefore, by definition, are not susceptible to a purely quantitative analysis [Kingdon, 97].
• financial systems are nonlinear and instationary. The vast number of financial market models derived from financial data by means of statistically based methods is linear. Necessary are tools that describe nonlinear instationary dynamic financial systems.
• using a wide spectrum of mathematical methods, many trading indicators have been developed. One important disadvantage of all indicators computed using historical data is: since historical data are used exclusively, the trading signal will probably lag advantageous trading points in time due to some necessary noise filtering of these data (in most cases averaging). This time delay may lead to significant losses. Assuming efficient markets, only predictive information can give some advantage here.

The objective of a self-organizing data mining driven financial trading system is to derive a trading signal from historical data using two kinds of models: A prediction model and a decision model. In a modeling /prediction step, self-organizing data mining is used to extract hidden knowledge from data fast, systematically and objectively. A decision model is responsible for signals generation based on the predictions provided by the prediction model. Such a predictive control is shown in figure 1.

Fig. 1: Predictive controlled financial trading system

 

In our simulations, we have used KnowledgeMiner's Analog Complexing algorithm to generate successively 5-day predictions of the evaluated stock, and the widely used MACD trading indicator (Moving Average Convergence Divergence) was chosen as a decision model. Here, however, the MACD was calculated using historical price data AND predicted prices correspondingly. This is equivalent to predicting the MACD 5 days ahead. For ideal predictions (zero error), this means that the indicator's time delay can be reduced by 3 days - ideally, a 3-day advantage relative to other market participants.

The question is, however, what a performance an almost real-world trading simulation can show? Our test was based on the following daily procedure. From a given set of historical daily price data of the NASDAQ index and a certain stock, the latter is predicted five days ahead using Analog Complexing. Then, the MACD is calculated on both historical prices and the predicted prices. The predicted MACD in turn is used to generate buy/hold/sell signals in the known way. If a trading action is suggested, the transaction is reserved to be executed at the next day's close price. No transaction cost was considered. When the market is closed, the new close prices are added to the data base and the procedure repeats the next day. This procedure installs moving modeling, and the performance results are based on out-of-sample predictions.

Results of Intel Corp., Novell, Inc., and Sun Microsystems (see also: [Lemke/Müller, 97] and [Müller/Lemke, 99]) show that the tested trading system of a 5-day predicted MACD indicator

• generates powerful, reliable 5-day predictions using Analog Complexing (in average above 95% accuracy);
• can reduce the indicator's time lag by 1-2 days;
• performs (much) better than a common MACD based trading system at same conditions (in average 25% profit gain).

These advantages can be explained exclusively by the inclusion of useful predictive information generated from data automatically by KnowledgeMiner.

However, a predictive controlled trading system cannot overcome a possible inherent weakness of the decision model. For the MACD, for example, it is that it may generate false signals in non growing/falling time phases. Here, a prediction will only have the effect that it generates a false signal one or two days earlier. Also, MACD is very sensitive on temporarily changing trends. So, decision model design needs some improvement.

 

References:

Kingdon, J.: Intelligent Systems and Financial Forecasting, Springer. London, Berlin, 1997

Lemke, F.; Müller, J.A.: Self-organizing Data Mining for a Portfolio Trading System. Journal of Computational Intelligence in Finance , 5(1997)3, pp.12-26

Müller, J.A.; Lemke, F.: Self-Organising Data Mining. Extracting Knowledge From Data. Libri, Hamburg, 2000