0
$\begingroup$

I am playing with RNNs / LTSMs for a classification task in predicting financial data.

I have a time-series going many years back, and are planning to divide it into a number of shorter time-intervals that will be used to obtain a classification.

For example divide the data into random length intervals from 30 minutes to 24 hours, with a binary response 'higher' and 'lower' that answers if the value goes higher or lower for the next hours after the interval.

My question is if I can "reuse" the data, as in for example for a certain 24 hour period create the following intervals: 24 hours, 12 hours, 6 hours, 3 hours and 1 hour and train a model on all of these examples. If I do so, obviously the same data will be reused with different responses.

Or is it better to create a number of models; for example a separate model for 24 hour intervals than the 3 hour intervals?

$\endgroup$
0
$\begingroup$

I dont see a problem in „reusing“ observations. Its just new/different info coming in. The question is more if the model will train/perform well with a short lookback or timelag. Why dont use a long(er) sequence of data to predict (eg) the movement in 3h steps. I wonder why you dont use all the data you have to train the LSTM and update the data stepwise. More data usually is better!

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.