


scikits.statsmodels.sandbox.tsa.kalmanf
---------------------------------------
ARMA  :      ARMA model using the exact Kalman Filter
StateSpaceModel  :
kalmanfilter  :      Returns the negative log-likelihood of y
conditional on the information set
kalmansmooth  :
updatematrices  :      TODO: change API, update names


scikits.statsmodels.sandbox.tsa.arima
-------------------------------------

runs ok, no refactoring bugs
has examples and monte carlo that can be split up into example files

ARIMA  :  currently ARMA only, no differencing used - no I
arma2ar  :  get the AR representation of an ARMA process
arma2ma  :  get the impulse response function (MA representation) for
ARMA process
arma_acf  :  theoretical autocovariance function of ARMA process
arma_acovf  :  theoretical autocovariance function of ARMA process
arma_generate_sample  :  generate an random sample of an ARMA process
arma_impulse_response  :  get the impulse response function (MA
representation) for ARMA process
arma_pacf  :  partial autocorrelation function of an ARMA process
deconvolve  :  Deconvolves divisor out of signal, division of
polynomials for n terms
index2lpol  :  expand coefficients to lag poly
lpol2index  :  remove zeros from lagpolynomial, squeezed
representation with index
mcarma22  :  run Monte Carlo for ARMA(2,2)


scikits.statsmodels.sandbox.tsa.varma
-------------------------------------

just filter experiments
needed to fix import for acf example

VAR  :   multivariate linear filter
VARMA  :   multivariate linear filter

scikits.statsmodels.sandbox.tsa.varma_tools
-------------------------------------------

Helper and filter functions for VAR and VARMA, and basic VAR class
needed import fix in top of module
maybe rename to varma_process
in "main" example for VarmaPoly, and some Var fit
Var could be used for Granger Causality tests, otherwise it's pretty limited

Var  :
    simultaneous OLS estimation
VarmaPoly  :  class to keep track of Varma polynomial format
    working with and transforming VARMA Lag-Polynomials (3d)
ar2full  :  make reduced lagpolynomial into a right side lagpoly array
ar2lhs  :  convert full (rhs) lagpolynomial into a reduced, left side
lagpoly array
padone  :  pad with zeros along one axis, currently only axis=0
trimone  :  trim number of array elements along one axis
varfilter  :  apply an autoregressive filter to a series x
vargenerate  :  generate an VAR process with errors u
varinversefilter  :  creates inverse ar filter (MA representation) recursively


scikits.statsmodels.sandbox.tsa.try_fi
--------------------------------------

(not included by script that generates this list)
various functions to build lag-polynomials for fractional and seasonal integration
and function ar2arma minimizes distance in terms of impulse response function

move these to a module or rename

scikits.statsmodels.sandbox.tsa.try_var_convolve.py
---------------------------------------------------

(not included by script that generates this list)
two functions:
arfilter : autoregressive filter for 1d, 2d and 3d
fftconvolve : multidimensional filtering using fft

many examples, but I'm not sure this (fft) is correct
incompletely copied for interpreter session
currently raises exception because a variable (imp) is not defined

scikits.statsmodels.sandbox.tsa.try_var_convolve.py
---------------------------------------------------

(not included by script that generates this list)
includes functions for
detrending,
(theoretical) acovf and similar for special cases
acf plot functions

(partially copied from matplotlib.mlab)

currently exception:  FIXED
uses arima.ARIMA class without data in constructor, and order now has 3 values and
is keyword with tuple as value

move plot function to new graphics directory ?


scikits.statsmodels.sandbox.regression.mle
------------------------------------------

one refactoring bug fixed, because arima.ARIMA needs data, use class method instead
runs without exception, but I didn't look at any results
"main" has quite a lot

AR  :      Notes
Arma  :      univariate Autoregressive Moving Average model
Garch  :  Garch model gjrgarch (t-garch)
Garch0  :  Garch model,
GarchX  :  Garch model,
LikelihoodModel  :      Likelihood model is a subclass of Model.
TSMLEModel  :      univariate time series model for estimation with
maximum likelihood
garchplot  :
generate_garch  :  simulate standard garch
generate_gjrgarch  :  simulate gjr garch process
generate_kindofgarch  :  simulate garch like process but not squared
errors in arma
gjrconvertparams  :      flat to matrix
loglike_GARCH11  :
miso_lfilter  :      use nd convolution to merge inputs,
normloglike  :
test_gjrgarch  :
test_misofilter  :

Other
-----
diffusion: continuous time processes, produce nice graphs but parameterization is
a bit inconsistent.


script files
============

sandbox/tsa/try_arma_more.py
----------------------------

imports scikits.talkbox which is not compiled against my current numpy and doesn't
run
contains
arma_periodogram : theoretical periodogram


Proposed Structure (preliminary)
================================

arima_estimation
----------------
ARIMA class for estimation, wrapper or containing different estimators
other wrappers: here or in separate ???
  - support for choosing lag-length


arma_process
------------
all theoretical properties for given parameters
simulation method with options: initial conditions, errors, (?) not sure what else

varma_process
-------------
including VarmaPoly and impulse response functions

filters
-------
miso_filter (should be in cython eventually)
ar_filter : fast VAR filter with convolution or fft convolution
(not sure what's the relationship between the two)
others ???

stattools
---------
empirical properties
acf, ...

tsatools
--------
helper functions
lagmat
detrend ???

others, unclear
---------------
???

open questions
==============

support for exog
----------------
is incomplete or missing from some implementations
not clear parameterization
- ARMAX  A(L)y_t = C(L)x_t + B(L)e_t
- ARMAX-simple  A(L)y_t = beta x_t + B(L)e_t
  Note: covers previous version by extending x_t

- ARMA residuals y_t = beta x_t + u_t, and A(L)u_t = B(L)e_t
- ARMAX 2-step A(L)(y_t - beta x_t) = B(L)e_t
  Note: looks the same as ARMA residuals, implies
      A(L)y_t = A(L)x_t + B(L)e_t

- ARMAX  A(L)(y_t - A^{-1}(L) C(L) x_t) = B(L)e_t
  this doesn't look useful, unless we cutoff A^{-1}(L)

problem: signal.lfilter can only handle ARMAX residuals model (I think)
deterministic trend have ARMAX-simple model, e.g. in unit root tests


support for seasonal and "sparse" lag-polynomials
-------------------------------------------------
- fit functions need support for different lag structures,
  e.g. zeros, multiplicative
- support for pre-filters, e.g. (seasonal) differencing





