Darren Winters

Are they any good? The simple answer is yes! Several scientific studies by leading universities in the United States such as Stanford, have shown that ANNs can give a distinct advantage over both experienced human stock traders and Expert Systems. In these studies, the neural networks are trained on past price changes in the stock market or currency market, and then their predictions are compared to subsequent prices to see if the advice they give is valid. It is not generally recorded whether the researchers have made any real money using them, possibly due to the ethical considerations that might arise from such speculation.

It won’t surprise you to learn that stock price prediction was one of the first uses of Artificial Intelligence, but in the early days, programmers tried to encode the sort of rules of thumb that human predictors used, such as “If the price of a share goes above a 7-point moving average, issue a buy signal.” Systems based on series of rules like this are called Expert Systems, and they have been used with some success. The main disadvantage that Expert Systems have compared to ANNs is that they require the rules to be explicitly programmed into them. These rules are often difficult to derive from stock predictors (many of whom seem to work on little more than instinct) and are subject to error and over-generalisation. ANNs, on the other hand, if given sufficient, typical training data, can derive the most suitable rules for themselves. On the other hand, ANNs are bad at describing those rules as they are stored in the form of numbers (weighted connection strengths). Expert systems can easily “explain” how they reached their conclusions.

One thorny issue with neural networks is that the inputs need to be coded properly in order that the network can extract useful information from them. For instance, training the network simply on the share prices themselves would be unsatisfactory, as that particular pattern of prices would depend on the absolute value of the stock market. In many cases, the obvious strategy is the correct one, and in the case of stock market prediction an obvious strategy suggests itself. It is not the absolute value of the stock prices that is of interest, but the changes in prices from one time period to the next. This is where the money is to be made, regardless of whether the general value of shares is hovering around the 4000 mark, the 5000 mark or the 6000 mark. For this reason, it would make more sense for the inputs to the ANN to be the changes in share price.

ANNs essentially associate input patterns with output patterns. The inputs could be the raw stock market data, since this is the material that technical analysts use to predict movements in the market. The outputs could be any one of several things. For instance, given inputs representing the share prices on day 1, day 2 and day 3, the output might be a prediction of the share prices on day 4 (or possibly even on days 4 and 5). Alternatively, the outputs might be simple “buy” and “sell” signals for shares in particular companies.

The advantage of ANNs is that they are flexible. They will, if correctly trained, learn to classify any pattern in the training library correctly: if they are given a pattern on which they have been previously trained, they will produce the correct output. However, they will also produce the correct output if they are given a pattern similar to one they have seen during training. They will therefore classify patterns that they have never seen before, based on the closest matching training pattern.

Whereas conventional computer programs are written specifically and involve specific rules given by the human programmer, ANNs are trained by example. They learn to associate patterns on their inputs with corresponding patterns on their outputs. Training the neural net involves giving it a sample from the library of known input patterns and the corresponding desired output. Then the neural network is told to adapt its connections. This is repeated with other input-output pattern pairs. Typically, the set of input-output pairs used for training the net has to be presented repeatedly many times, and so neural networks can often take several hours to train. Once the neural network has been trained, you can present an input pattern to it and it will produce the corresponding output pattern.

An ANN consists of a number of neurons simulated in a computer with connections between them. Each of these connections has a certain strength, or “weight”, indicating how strong the connection is. A connection with a large weight is efficient at passing on a signal from one neuron to another. The information and expertise that ANNs possess is coded in these strengths between neurons.

Simulated neurons can be connected in various patterns, termed “architectures”. Although complex architectures exist (for instance in Cell Assemblies, which model closely structures in the human brain), the simplest architecture, the Multi-Layer Perceptron, is probably the most common. An MLP consists of neurons in distinct layers, each of which is connected to the next layer along. Signals can pass only one way through the MLP, from the inputs of the first layer to the outputs of the last one.