First charts

27Jan10

The implementation was done in Matlab. The Neural Network inputs are
the previous 5 days volume and close price of the SPX.
It used 200 neurons an 1 layer. The training is happened from 2000-01-01.
To illustrate why NN prediction is not so easy, I insert here a couple of charts:

These charts represents the %gain of the strategy playing long or short (no neutral allowed) at the end of the market for 100 days.
100 days is 5 months.
The annual (260days) %gain values for the different runs were: 2.23%, 20.86%, -36.08%, 8.95%, 34.18%.
The average of the 5 is 5.8% annual. Before the commission cost.
Considering the commission cost it is not good.
And the scary thing is the -36% %gain from the 5 runs.
So, currently, I consider that the gain (if any) is not significant.
More research is needed.
I reckon the current result is totally random.
I should decrease the Neuron numbers from 200 to about 50 to have better generalization of the NN.

Filed under: Uncategorized   |  Leave a Comment