You could try setting the covariance matrix type to 'tied'. This will force a
single covariance matrix across all states.
--
Helge Reikeras
On Monday 26 August 2013 at 4:03 AM, Shuo Wang wrote:
> Hi Helge,
> thanks for the reply.
>
> I have tried using first derivative to fit hmm. This is the results
>
> 0th hidden state
> ('mean = ', array([ 0.08534182, 0.07491009]))
> ('var = ', array([ 1.45029737, 1.23234594]))
> 1th hidden state
> ('mean = ', array([-0.25764837, -0.24770928]))
> ('var = ', array([ 0.06077981, 0.06085699]))
> 2th hidden state
> ('mean = ', array([ 0.10585516, 0.10645033]))
> ('var = ', array([ 0.04375404, 0.04436854]))
>
> The problem is that hmm seems to separate the states by variance, because in
> 0th state, I have big rise and down portions. This is similar to the stock
> example here
> http://scikit-learn.org/stable/auto_examples/applications/plot_hmm_stock_analysis.html#example-applications-plot-hmm-stock-analysis-py
>
>
> The 4th state identified is based on variance also. So how can I force the
> state detection based on mean instead?
>
>
>
> On Mon, Aug 26, 2013 at 1:31 AM, Helge Reikeras <[email protected]
> (mailto:[email protected])> wrote:
> > Hi
> >
> > 1. From the source code it looks like predict is just a wrapper around
> > decode. I see predict only returns the state sequence with largest
> > likelihood while decode also returns the actual value of the log
> > likelihood.
> >
> > 2. I'm no expert on modelling financial data but it sounds like you are
> > mostly interested in the price change so using the first derivative (and
> > maybe 2nd and higher derivatives) seems reasonable. As I guess the data is
> > noisy you should probably compute these using something like SavitzkyGolay
> > filter:
> >
> > http://wiki.scipy.org/Cookbook/SavitzkyGolay
> >
> > Then the sign of the value of the first derivate feature in the state
> > distribution mean will tell you if the state represent "up" or "down"
> > (apply a distance from 0 threshold for "neutral") and the absolute value
> > will give you the magnitude.
> >
> > Regards,
> >
> > --
> > Helge Reikeras
> >
> >
> > On Sunday 25 August 2013 at 4:41 PM, Shuo Wang wrote:
> >
> >
> >
> > > Hi,
> > >
> > > I have two questions regarding the hmm model from a user's perspective,
> > >
> > > 1. what is the difference between method decode and predict, they seems
> > > to be having the same description
> > >
> > > 2. If i fit a financial time series into hmm, hmm seems to automatically
> > > detect states based on mean and variance, so if I want it to detect
> > > states related to regimes(up, down and neutral trends), I have to
> > > transform price into binary time series, such as 1 indicates current
> > > minute is a rise compare to previous minutes, -1 indicates current minute
> > > is a drop compare to previous minutes, 0 is unchanged. But if I transform
> > > my price info into binary information, I loss magnitude information. How
> > > can I keep the magnitude information, yet still let the hmm detect up,
> > > down and neutral trend, instead of mean and variance regimes.
> > >
> > > --
> > > 王硕
> > > 邮箱：[email protected] (mailto:[email protected])
> > > Whatever your journey, keep walking.
> > > ------------------------------------------------------------------------------
> > > Introducing Performance Central, a new site from SourceForge and
> > > AppDynamics. Performance Central is your source for news, insights,
> > > analysis and resources for efficient Application Performance Management.
> > > Visit us today!
> > > http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
> > >
> > > _______________________________________________
> > > Scikit-learn-general mailing list
> > > [email protected]
> > > (mailto:[email protected])
> > > https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
> > >
> > >
> > >
> >
> >
> >
> > ------------------------------------------------------------------------------
> > Introducing Performance Central, a new site from SourceForge and
> > AppDynamics. Performance Central is your source for news, insights,
> > analysis and resources for efficient Application Performance Management.
> > Visit us today!
> > http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
> > _______________________________________________
> > Scikit-learn-general mailing list
> > [email protected]
> > (mailto:[email protected])
> > https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
> >
>
>
>
> --
> 王硕
> 邮箱：[email protected] (mailto:[email protected])
> Whatever your journey, keep walking.
> ------------------------------------------------------------------------------
> Introducing Performance Central, a new site from SourceForge and
> AppDynamics. Performance Central is your source for news, insights,
> analysis and resources for efficient Application Performance Management.
> Visit us today!
> http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
>
> _______________________________________________
> Scikit-learn-general mailing list
> [email protected]
> (mailto:[email protected])
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>
>
------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
