Hi I have been using Subutai's sine examples as a basis to do some research on financial trends. Thanks to Subutai for a great tutorial, it was quite easy to follow.
My data consists of a number of years of tic data for a number of currencies. I have discovered quite quickly that with a single hierarchy that tic data is nigh on useless, so the data has to be aggregated first. To do this I have used standard 60 minute 'candles' - each 60 minute candle has period open, close, high and low. For better modelling you *need* the variance as this "noise" doesn't represent errors as in audio processing, it is actual prices and if you were to trade the noise can make a significant difference to the outcome of the trade. Helpfully I have found that by using open, high and low you get better predictions than over merely using open. Would a simple high - low variance instead make sense? What I want to do now is to use mix in other currencies to see how that affects the results. It strikes me that it would be important to tell nupic that different columns should be grouped together ie "open, high, low for currency a" would be one group, etc. This would be particularly important if I was to use a variance field for each currency instead of high and low. How would I go about doing this? Thanks Mark Mark Wilson
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
