Neural Network
Human brain has approximately 10^11 neuron and 10^15 connections amongst neurons
building a huge neural network. Artificial Neural Network (ANN) is built to model part
of the human brain. Its neuron is represented by memory cell and the connection between
cells is described by a weight value. Its output is calculated using the wights of the
connections and the input values. Training is a process to adjust the wights and
connections. The network is present with the input and the desired output. The wights
and connections are adjusted so that the output is produced. Once a network is trained
correctly, it discovers the rules between the input and output rather than memorizing
them. The network will be able to give outputs from a different set of input. In a sense,
the network has become an expert for this particular type of input.
Unlike human brains which are emotional and have hard time discovering hidden rules,
Neural Network is well suited for financial application. In addition, human adapts fast
changing environment slowly, whereas Neural Network can adapt to new rules just by
retraining. Numerous sophisticated financial institutes have been reported using Neural
Network to make trade decisions. The process of training a network involves data gathering,
input/output manipulation and normalization. Furthermore, Since neural network training requires
relatively high amount of computing power and developing a good network needs a lots of resource, it was only available to large organization and rich clients. The recent advances in personal computer make it possible for general users to perform such operation. Trade Method's artificial intelligence capabilities are result of years of dedicated research and development in Neural Network and Genetic Algorithm. It allows users to make better
trade decisions based on this cutting edge technologies. Users can easily create an Neural
Network model that adapts to the current market condition without knowing the
complexities of AI.
Trade Method gives user the ability to build a model, train it and execute it as an indicator. User can build nested Network by using the trained indicator as an input. There are three types of learning; Turning Point, Probability and Trend forecast.
The following is screen capture of Trade Method's Neural Network training screen:
Genetic Algorithm
Genetic algorithms (GA) are inspired by Darwin's theory of evolution by natural selection. It solves problem and optimize solutions by continuously evolve the solutions. In biology, chromosomes are strings of DNA that carries genetic information. GA represents a solution to a problem using a string of numbers similar to chromosomes. One generation or population consist of a set of chromosomes. It use a process similar to Natural Selection by which chromosomes with favorable characteristics are carried over to the next generation. New generation of chromosomes are created using mutation and/or crossover. Mutation is a way one or more gene values in the chromosomes changes, its goal is to introduce new genes in the new generation. Crossover is another operation to combine two chromosomes primarily from the previously generation (parents) with hope that the new chromosome may be better than both of the parents A function is used to measure the fitness of the solutions. As a result of this evolutionary process, a good solution is found.
Trade Method use Genetic Algorithm to select the right inputs for a neural network. It evolves the neural network so that an optimized neural network is found. The fitness function to evaluate the network is a financial performance function to measure the accuracy of the neural network signal. Traditional human process to select a good neural network is to continue to repeat the training and testing procedure until a network is found. This process can be very time-consuming and the search space grows as the number of parameters increase. With Trade Method's Neural Genetic Search, all these processes are made easy.
To start the search process, users need to specify the dates for the network to learn. In order to find a generalized network, It is recommended to select at least two years of data. In addition, users should choose data with both oscillation and trend as well as bull and bear market. Walk-forward testing is a methodology that is used to ensure the trained network is producing result correctly. Users choose an interval of time to test the network. The time interval does not have to be after the learning end date. There are three types of forecast users can choose from. Turning Point forecast give signal when a security is changing directions. Probability Forecast predicts the chance a security is going to move up or down. Trend Forecast predicts whether the coming trend is bullish or bearish. Finally, the user choose the model to optimize and the number of inputs to select from the model. For example, a Model can contain 40 indicators, the number of inputs can set to 10.
Forcasting Signals
Trade Method offers three types of forecast users to choose from.
Turning Point Forecast give signal when it predicts a security will changing direction.
When the predicted value is close to 1, the direction will likely to change to reverse from
a up trend. On the other hand, when the predicted value is close to -1, the direction will
likely to change to an up trend. Combining the Turning Point Forecast with Trend Forecast
will give user a clearer picture.
The following is a screen capture of QQQQ from August 2004 to June 2005. The bottom
panel shows a plot of the neural network turning point signal. A signal with a value
greater than 0.3 indicates sell (B,D,F) and a signal with a value less than 0.3 indicates buy
(A,C,E,G). Users can use more stringent values for their signal.
Probability Forecast predicts the chance a security is going to move up or down. It more
appropriate to use the signal for trading in the shorter term.
Trend Forecast predicts the whether the coming trend is bullish or bearish. It continuously
generate signal for the trend strength
The signal produced by the network can be combined with other indicators to create a complete trading system.