|
PureBytes Links
Trading Reference Links
|
From: "Frode L. Aschim" <faschim@xxxxxxxxxxxx>
>>However, even though the predicted output was very good, when I added new
data that the network hadn't "seen" before, the predictions were much less
accurate. There are a few ways to combat this, but it made me think if
neural networks would work in any market as the conditions change all the
time. <<
Very good issues you brought up.
1. You can always expect a model to perform less well on new data. The mark
of a reliable model is how little the performance degrades. If your model
degraded a lot, then its no good. Try again.
2. The nonstationary aspect of the market means your model's assumptions about
the market's input-output statistical relationships may eventually become
invalid. So you will need to perform periodic retraining.
3. There is the old question -- how much data to use? If too little, then the
model will memorize special cases and be useless when forecasting. If too
much, then the model will learn input-output relationships that are no longer
playing out in the market, again degrading model performance. There are ways
to mitigate this problem somewhat, but it'll never go away.
- Mark Jurik
********************************************
Tools for financial data preprocessing
Jurik Research http://www.jurikres.com/
********************************************
|