An Approach to Model Development

Stock Selection

Some stocks have more predictable trading patterns than others and are thus easier to model. TradingSolutions makes it relatively easy to find these predictable stocks by giving you the ability to train a neural network for each stock in a group, then sort the stocks in the group by the profitability of the resulting trading model.

The process of selecting a stock starts by loading in the data for a large group of stocks, such as the S&P 500 or the NASDAQ 100, into the TradingSolutions portfolio. A neural network prediction is defined for the entire group and the training stage is initiated. After a couple of hours of processing (the time is dependant on the number of stocks in the group and the speed of the computer), TradingSolutions sorts the stocks in the portfolio by the profit of the underlying models. An example of this sorted portfolio is shown in Figure 1.


Figure 1: Performance summary of  S&P 500 models, sorted by profit.

Input Selection

TradingSolutions includes over 200 pre-defined indicators that can be used as inputs into a neural network. In addition, the software allows you to define your own indicators with its unique formula building system. Somewhere in the neighborhood of 3 to 10 data fields and indicators are selected as inputs to the neural networks that are applied to the group of stocks. Most of these inputs ones that are often used in traditional technical analysis, such as moving average, stochastic oscillator, MACD and RSI. 

Desired Output

The traditional approach to developing a neural network trading model is to first train the neural network to predict the value of the closing price one or more days into the future, then define an entry/exit system based on that prediction. TradingSolutions offers an alternative technique known as "optimal signal modeling".

This revolutionary approach first determines the best action based on future prices and uses this information to produce an optimal signal. As an example, if the price of a stock were to rise 10% from January 1 to January 10, then the optimal signal for January 1 would likely be a value in the "buy" or "enter long" range. A neural network could then be trained to produce a "buy" signal on January 1, as well as any other dates that have similar input characteristics. By setting the desired output of the neural network to be the optimal signal, the neural network is able to learn the relationships between the network inputs and the corresponding buying and selling opportunities.

Data Set Selection

The more data you can use to train a neural network the better your results are likely to be, provided that the data is applicable to the current time frame. For most stocks this optimal time frame for training data is the preceding 3-5 years. When creating real-time models with intraday data, a much shorter time frame can be used since each day contains more data.

Neural Network Model Selection

TradingSolutions has over 50 neural network architectures to choose from. Many of the best performing TradingSolutions models developed to date have belonged to the class of neural networks known as "Time-Lagged Recurrent Networks" or TLRNs for short. TLRNs use a sophisticated learning algorithm known as backpropagation through time. This algorithm trains over a trajectory of the input space, enabling it to capture the temporal dynamics of the signals. 

Trading the Model

Once you have a model that has proven to be profitable, setting up TradingSolutions to give you the daily entry/exit (buy/sell) signals is very straightforward. Figure 2 shows a sample portfolio of stocks, each of which has a neural network model associated with it. This example shows two groups of stocks -- one that is currently being traded and one that is being watched for future trading opportunities.

Figure 2: TradingSolutions portfolio with entry/exit signals for the next day.

At the end of each trading day, you should first obtain the data for the stocks in your portfolio. This can be done automatically using a direct connection to the ProphetFinance.com or eSignal data services via the internet, or you may import the data from text files obtained from a variety of sources. 

Immediately after the download/import, TradingSolutions will perform the necessary calculations needed to generate the entry/exit signals for the following day. The first column of Figure 2 shows this signal for each stock in the portfolio. Those signals that have a pink background indicating that there is a change in position. For example, the signal for "Apollo Group" shows a solid red triangle indicating that a short position should be taken the next day. If the current position is long, then there would be two trades to be make -- one to exit (sell) the long position and one to enter a new short position.

The second column shows the closing price for the day. The third column is the estimated return on investment if you had been trading this model for the past 3 months. The last column shows the estimated return for the past 6 months.