View Single Post
  #2  
Old 07-04-2012, 02:36 PM
htlin's Avatar
htlin htlin is offline
NTU
 
Join Date: Aug 2009
Location: Taipei, Taiwan
Posts: 601
Default Re: Machine Learning

Quote:
Originally Posted by dbl001 View Post
Say I am applying machine learning methods to forecast the future price movement of the Euro. I have 10 years of data for training, testing, etc.

We have learned that future events which will dramatically affect the prediction (sovereign defaults,
country leaving the EU, multiple countries leaving the EU, etc.) have never occurred, were viewed as impossible months ago, and have no basis in the EU treaties or ECB regulations and mandates.

Would applying machine learning to this data be considered as having 'false assumptions'?

Are there methods for determining when this is happening when applying machine learning to a dataset, and quantifying the reliability of the predictions, or is this a matter of judgement left to the practitioner?

Thanks in Advance!
In general, ML is not a magic box and cannot magically guess what would happen in the future, unless the future contains some "predictable" nature from the current data. The unpredictable parts acts like the "deterministic" (or possibly stochastic?) noise in this learning model, which is not possible to fit and can cause the overfitting problem. Following the suggestions within the book, what you can do is to try to "stay safe" with the low-complexity models and regularization. Hope this helps.
__________________
When one teaches, two learn.
Reply With Quote