Designing A Personal Neural Net Trading System by James Stakelum
What goes through the mind of someone designing a neural network trading system? Last month, James Stakelum explained some of the steps involved in putting together a neural network. This time, he talks about the choices he made to design a neural network to apply to the financial markets specifically.
Last time, I detailed the general issues I found important regarding the overall design and composition of creating my own neural network. This time, I want to explain the choices I made to design a neural network for the specific application of trading financial markets. I used a historical database of the major futures contract markets with the date, open, high, low and close to train and test my neural networks. I had to overcome one obstacle: I wanted more granularity of prices than provided by the daily four prices in the database. An ideal series for my purposes would have been the actual price ticks for each day, but such databases are prohibitively expensive to purchase and fortunately were not absolutely necessary.
I had to produce a more realistic flow of price data than the daily database provided. To do so, I transformed the four daily prices into a simulated tick database by generating a series of prices for each day that fluidly connects the open and close via the shortest path between the high and low (Figure 1). The tick data was only used in testing the network, when trading simulations are run testing hypothetical buys and sells. Daily data with only date, open, high, low and close, not tick data, was used for training.