|
PureBytes Links
Trading Reference Links
|
On Sun, 12 Jul 1998 Peter2150@xxxxxxx wrote:
> The more correlated the data, the less valid the test. What you should do is
> train on the 1990 - 1995 data, and then test on data from 1996 to present.
> However no matter how good that may look, if the market changes character in
> the future, you still have garbage. I wish you luck.
>
But this is what I wanted to avoid by using a different market. I just
don't have enough data to train AND test. 1990 ~ 1995 train is good, but
I just don't feel comfortable with testing data of only 1996 ~ 1997 and
half of 1998. I could go back further with training but the market has
structurally changed and I don't believe it relevant to use them as
training sets.
|