Tobias Hahn - Designing Stock Market Trading Systems: With and without soft computing

• Twitter
• Facebook
• Google Plus

All trading systems have some behaviour which we would like to modify. Finding variables to help us influence this behaviour, however, is often very difficult.

Consider the case of a simple moving average crossover system. Such a system will invariably be subject to whipsaw. This is where a signal occurs, but is quickly reversed, maybe even in the next time period. This causes the system to move in and out of trades, tying up capital, and incurring transaction costs. The typical traders solution is to implement some crossover threshold which must be exceeded before the trade is taken, but this delays entry, and in the case of a profitable outcome, costs time and money.

A better solution might be to look to a neural network to help solve this problem. An ANN is a universal approximator; it is capable of determining a non-linear relationship between a set of inputs and a set of outputs.

The main challenge is to find a way to express your problem so that an ANN can help provide the solution. Once this is done, the process of creating the required ANN can begin.

The main functions in building an ANN are:

1. Expressing the problem to be solved in such a way that an ANN can be used to build the solution
2. Partitioning data into training and testing sets
3. Finding variables of influence
4. Making ANN architecture choices
5. Training
6. Testing

Whilst ANNs can help solve a great many complex problems, it should be remembered that there are a number of limitations inherent in the use of neural networks. Chiefly, these concern the fact that the neural net is a black box, and that rule extraction, whilst possible in some limited circumstances, is a particularly difficult and uncertain process. Overall, however, the neural net is not suitable for use as an explanatory tool.

Neural networks also tend to overfit the data if not very carefully controlled during the training process, and can find non-causal patterns in data very easily. There are no rigorous training methodologies that avoid this problem entirely. Determining a good internal structure for the network also tends to be a rather delicate process and although a number of useful guidelines exist, there are again no definite steps to success.

Despite these clear limitations with neural networks, they are still considered the tool of choice for investigating non-linear relationships amongst noisy and complex data sets.