All,
Just wanted to post my findings after playing with a few options in NS. I have tried building a few new nets based on the work of Len and Jay. I use the stock (20 to 30 s/t, l/t, price and volume at 1), 1 related sector index (1 l/t), and a market index (1 l/t). Then I set the number of neurons to between 8 and 12. I also set my training period to be less than a year, and my verification to be 2 to 3 months.
What I have found is that nets train extremely quick, often in 12 hours or less. Profitability is very good also, both in sample and out of sample. I think by reducing the number of neurons you force NS to generalize on the training data more, and by using a shorter training period you pick up only the most recent patterns in market. While I still have some investigating to do, NS has managed to give some pretty good signals (NN two days ago, up about 10% right now).
Maybe some others could try this and post their findings?
As a side note, the Genetic process also converges much much quicker this way. By reducing the number of variables, you reduce the search space.
Optim |
