Thank you, i will review it tonight. On Wed, Dec 3, 2014 at 3:57 PM, Chetan Surpur <csur...@numenta.org> wrote:
> You may find this video helpful in understanding the characteristics of > applications that are good fits for HTM: > http://numenta.com/learn/applications-of-hierarchical-temporal-memory.html > > > On Dec 2, 2014, at 6:53 PM, Daniel Bell <john.mrdaniel.b...@googlemail.com> > wrote: > > How do we know if a problem is intractable for nupic? Are there any > guidelines or rules? > > On Wed, Dec 3, 2014 at 3:51 PM, Dennis S. Sedov <st...@acortix.com> wrote: > >> I have been thinking about this problem and came to the following >> conclusions: >> >> 1) Input space is very large if we want to obtain correct predictions >> most of the time. As Matt correctly noted - there are too many variables. >> Besides, there are to many variables that we don’t even know about. Also, >> meaning of the variables is important. Some of the variables are not >> linear, and some variables are byproduct of other variables. There are just >> too many levels of data. Distinguishing what is data and what is not is a >> problem on its own. >> >> 2) It is, I assume, possible to predict the data “sometimes”. There are >> well known patterns in the stock market over a short period of time. Those >> are already being predicted by other financial analytical systems. >> >> 3) Another problem is the fact that by predicting market price and taking >> actions on that knowledge you change the market. This become a much bigger >> problem of predicting changes due to predictions. It’s a closed loop >> system. Any time someone comes up with a good model it spreads like a virus >> and stales very quickly. Again you have to take into account what you’re >> doing to the market by predicting its value. >> >> >> >> Sincerely, >> Dennis S. Sedov >> >> On Dec 2, 2014, at 6:43 PM, Daniel Bell < >> john.mrdaniel.b...@googlemail.com> wrote: >> >> That is certainly understandable and fair. So this is a practical >> limitation of not having visibility on all of the relevant factors. >> >> Could nupic do this if we theoretically did have all the features that >> represent the state of the system? >> Would a subset of these features, no matter how large, be able to resolve >> 'reasonable' predictions? >> >> >> On Wed, Dec 3, 2014 at 3:16 PM, Matthew Taylor <m...@numenta.org> wrote: >> >>> Hi Daniel, >>> >>> Can any one human being predict stock market prices with any accuracy? >>> If you think about how many factors actually affect even a single >>> stock price (economy, inflation, weather, time of year, time of day, >>> moods of investors, CEO scandals, other stock prices, I could go on >>> and on...), it would be extremely hard to identify them all, much less >>> isolate them into individual scalar values and feed them into NuPIC. >>> There are just too many unknown factors involved. Even the best human >>> minds can't do it. >>> >>> --------- >>> Matt Taylor >>> OS Community Flag-Bearer >>> Numenta >>> >>> >>> On Tue, Dec 2, 2014 at 5:51 PM, Daniel Bell >>> <john.mrdaniel.b...@googlemail.com> wrote: >>> > Hello, >>> > >>> > In one of the talks Jeff Hawkins mentioned that stock market data >>> cannot be >>> > predicted with numenta. Why is this the case? Is it not an appropriate >>> > problem space? >>> > >>> > My question here really is, what are the limitations and how do we >>> identify >>> > problem spaces that will work well with numenta and not work well >>> prior to >>> > an attempts to train/predict? >>> > >>> > Regards, >>> > >>> > Daniel >>> >>> >> >> > >
