Prev: why RMSE obtained from stepwisefit and regstats differs?
Next: The number provided for each value
From: emceeaye on 8 May 2010 04:52
Hi Matlab experts,
My background is in functional magnetic resonance imaging time series
data analysis, and I recently realized that there's no reason l can't
also apply the same statistical procedures I use for my research to
stock forecasting. I want to analyze historical time series data for
stocks that I'd like to import into Matlab? I have yet to use Matlab,
and so this project will be a good chance for me to get my feet wet.
For the time-series data, I plan to look at different sampling rates
(e.g., closing prices for the day; price 4 times a day, etc.) for
different ranges of time (e.g., past 1/2 year, past 1 year, past 2
years, past 5 years, etc.).

It would be great to get insights from people with experience with
time series data about the pre-processing steps necessary to analyze
this data. Traditionally for fMRI data, there are a series of pre-
processing steps researchers go through before analyzing different
time series of signal values (e.g., remove linear trend, smooth the
data, etc.) in order to free the data of contaminating artifact and to
make it safe to analyze (which will in turn make the statistics that
are based on them interpretable). Can I expect that the assumptions
that must be made about stock time series data before safe analysis is
possible are the same as for fMRI signal fluctuations time series data
(e.g. normal distribution, homogeneity of variance, etc.)? If so, do
they require specific pre-processing steps to prepare it for analysis
(detrending, removing outliers, etc.), and are these procedures
available in Matlab? I also wonder if there is an uncomplicated (or
as uncomplicated as possible) book that can do a good job of answering
these questions that anyone can recommend?

I would really appreciate any help with this as it will go a long way
for me.

Thank you in advance.
From: Rune Allnor on 8 May 2010 07:16
On 8 Mai, 10:52, emceeaye <dumathechee...(a)yahoo.com> wrote:
> Hi Matlab experts,
> My background is in functional magnetic resonance imaging time series
> data analysis, and I recently realized that there's no reason l can't
> also apply the same statistical procedures I use for my research to
> stock forecasting.

Yes, there is.

MRI is based on very specific physical relationships that can
be expressed compactly and conveniently by means of maths.

Stocks indexes / prices are random series that rely on unpredictable,
unforeseen factors, including human psychological factors like fear,
nervousness and mass hysteria.

Within the past couple of months Europe have experienced a couple
of factors that were unforeseeable, like the Iceland volcano
eruption and the Greek economical crisis. There is no way any
stock market model can include or account for such factors.

Rune
From: Wayne King on 8 May 2010 07:43
emceeaye <dumathecheetah(a)yahoo.com> wrote in message <3dfbec32-85bf-4bb5-a01a-b3479188e247(a)g1g2000pro.googlegroups.com>...
> Hi Matlab experts,
> My background is in functional magnetic resonance imaging time series
> data analysis, and I recently realized that there's no reason l can't
> also apply the same statistical procedures I use for my research to
> stock forecasting. I want to analyze historical time series data for
> stocks that I'd like to import into Matlab? I have yet to use Matlab,
> and so this project will be a good chance for me to get my feet wet.
> For the time-series data, I plan to look at different sampling rates
> (e.g., closing prices for the day; price 4 times a day, etc.) for
> different ranges of time (e.g., past 1/2 year, past 1 year, past 2
> years, past 5 years, etc.).
>
> It would be great to get insights from people with experience with
> time series data about the pre-processing steps necessary to analyze
> this data. Traditionally for fMRI data, there are a series of pre-
> processing steps researchers go through before analyzing different
> time series of signal values (e.g., remove linear trend, smooth the
> data, etc.) in order to free the data of contaminating artifact and to
> make it safe to analyze (which will in turn make the statistics that
> are based on them interpretable). Can I expect that the assumptions
> that must be made about stock time series data before safe analysis is
> possible are the same as for fMRI signal fluctuations time series data
> (e.g. normal distribution, homogeneity of variance, etc.)? If so, do
> they require specific pre-processing steps to prepare it for analysis
> (detrending, removing outliers, etc.), and are these procedures
> available in Matlab? I also wonder if there is an uncomplicated (or
> as uncomplicated as possible) book that can do a good job of answering
> these questions that anyone can recommend?
>
> I would really appreciate any help with this as it will go a long way
> for me.
>
> Thank you in advance.

Hi, I think you will find that all time series analysis problems share some similarities, but certainly each domain has some unique elements and problems as well. Specifically, the "homogeneity of variance" assumption that you referred to is violated consistently in econometric time series. As Rune points out, there are often shocks in financial time series.

MATLAB provides a lot of tools in the Econometrics Toolbox for modeling conditional heteroskedasticity like that encountered regularly in financial time series. Start with Hamilton's textbook on Time Series Analysis. You will find a lot of things that are familiar from working on fMRI time series and a lot that is probably not familiar.

Wayne
From: matlaberboy on 8 May 2010 13:45
James Hamilton's Time Series Analysis is not *simple/uncomplicated* (for most of us)!!!!!

I can highly recommend the following as being good applied examples:

http://www.amazon.co.uk/Introductory-Econometrics-Finance-Chris-Brooks/dp/052179367X/ref=sr_1_1?ie=UTF8&s=books&qid=1273340487&sr=1-1

http://www.amazon.co.uk/Market-Models-Guide-Financial-Analysis/dp/0471899755/ref=sr_1_1?ie=UTF8&s=books&qid=1273340516&sr=1-1
From: emceeaye on 8 May 2010 16:38
On May 8, 4:16 am, Rune Allnor <all...(a)tele.ntnu.no> wrote:
> On 8 Mai, 10:52, emceeaye <dumathechee...(a)yahoo.com> wrote:
>
> > Hi Matlab experts,
> > My background is in functional magnetic resonance imaging time series
> > data analysis, and I recently realized that there's no reason l can't
> > also apply the same statistical procedures I use for my research to
> > stock forecasting.
>
> Yes, there is.
>
> MRI is based on very specific physical relationships that can
> be expressed compactly and conveniently by means of maths.
>
> Stocks indexes / prices are random series that rely on unpredictable,
> unforeseen factors, including human psychological factors like fear,
> nervousness and mass hysteria.
>
> Within the past couple of months Europe have experienced a couple
> of factors that were unforeseeable, like the Iceland volcano
> eruption and the Greek economical crisis. There is no way any
> stock market model can include or account for such factors.
>
> Rune

Hi Rune,
Part of what you say is true and part is not. 1) While stock prices
are influenced by "unpredictable" and "unforseen" fluctuations in
values such as those seen in the recent volatility in stock market
prices and the other examples you gave, the prices certainly don't
"RELY" on them as you stated--after all, they are the exceptions
rather than the rules.

2) Certainly evidence suggests that fMRI data is based on biological
(or "physical" relationships) as you suggest, but there are a lot of
unknowns about the physiological significance of the BOLD signal
(e.g., is it blood flow, oxygen consumption, glucose consumption, a
combination of all, or a consequence of magnetic field
inhomogeneities, and the list goes on). Furthermore, the signal
changes being analyzed often do not reflect the underlying
psychological/cognitive process that the researcher is intending to
measure and the conclusions drawn and interpretations made from the
results of analyzing this data are often leaps in logic and not
necessarily accurate (happens all the time). Also, Just as human
"nervousness", and "self-consciousness" may influence stock prices,
these same types of unpredictable, uncontrollable and unverifiable
emotions/states and traits of the subjects may contaminate the sampled
BOLD signal as well, which ultimately interferes with the
intepretability of the data (happens all the time). so results of fMRI
data may be expressed "conveniently" and "compactly" by math, but what
is being deduced or extrapolated from the results is often far from
"convenient" or "compact".

Routine techniques are often applied to fMRI BOLD data to reduce
"noise" and "outliers" because they don't fit the expected behavior of
the signal--why is this done? to eliminate the influence of anomalies
and unexpected behavior in the signal being sampled (basically because
the researcher cannot explain the reason for them and so discards the
raw data for them). These same procedures of dealing with outliers and
noise may also be applied to stock time series data to eliminate their
influence on what the researcher is trying to determine by analyzing
the data. Does this mean that stock prices time series don't exhibit
more unpredictable volatility than fMRI BOLD data? Maybe or maybe
not. That is one of the questions I would like to answer.

One way to compensate for these unexpected fluctuations in the stock
market is to increase the sampling period (e.g., from 1 year to 2, 3,
4 years worth of values)--by increasing the sampling period you reduce
the impact of sudden unexpected fluctuations on the overall
statistical results. Either way, I don't have to know much about
econometrics to say that stock prices are certainly not "random" as
you stated. However, there are certainly many factors influencing the
values and it's trying to account for as many of these factors as
possible and determine their relative contribution to the results that
may be a bit of a challenge.

emceeaye
 |  Next  |  Last
Pages: 1 2 3
Prev: why RMSE obtained from stepwisefit and regstats differs?
Next: The number provided for each value