6.11.4. statsmodels.sandbox.tsa.fftarma¶
Created on Mon Dec 14 19:53:25 2009
Author: josef-pktd
generate arma sample using fft with all the lfilter it looks slow to get the ma representation first
apply arma filter (in ar representation) to time series to get white noise but seems slow to be useful for fast estimation for nobs=10000
- change/check: instead of using marep, use fft-transform of ar and ma
separately, use ratio check theory is correct and example works DONE : feels much faster than lfilter -> use for estimation of ARMA -> use pade (scipy.misc) approximation to get starting polynomial
from autocorrelation (is autocorrelation of AR(p) related to marep?) check if pade is fast, not for larger arrays ? maybe pade doesn’t do the right thing for this, not tried yet scipy.pade([ 1. , 0.6, 0.25, 0.125, 0.0625, 0.1],2) raises LinAlgError: singular matrix also doesn’t have roots inside unit circle ??-> even without initialization, it might be fast for estimation -> how do I enforce stationarity and invertibility,
need helper function
get function drop imag if close to zero from numpy/scipy source, where?
6.11.4.2. Classes¶
ArmaFft(ar, ma, n) |
fft tools for arma processes |
ArmaProcess(ar, ma[, nobs]) |
Represent an ARMA process for given lag-polynomials |
OneTimeProperty(func) |
A descriptor to make special properties that become normal attributes. |