carma_pack is an MCMC sampler for performing Bayesian inference on continuous time autoregressive moving average models. These models may be used to model time series with irregular sampling. The MCMC sampler utilizes an adaptive Metropolis algorithm combined with parallel tempering. Further details are given in this paper.
For a guided tour of carma_pack click here
or see the ipython notebook available under the examples/ folder.
To illustrate the important functionality of carma_pack, start by defining the power spectrum parameters for a CARMA(5,3) process:
>>> import carmcmc as cm
>>> import numpy as np
>>>
>>> sigmay = 2.3 # dispersion in the time series
>>> p = 5 # order of the AR polynomial
>>> mu = 17.0 # mean of the time series
>>> qpo_width = np.array([1.0/100.0, 1.0/300.0, 1.0/200.0]) # widths of of Lorentzian components
>>> qpo_cent = np.array([1.0/5.0, 1.0/25.0]) # centroids of Lorentzian components
>>> ar_roots = cm.get_ar_roots(qpo_width, qpo_cent)
>>> ma_coefs = np.zeros(p)
>>> ma_coefs[0] = 1.0
>>> ma_coefs[1] = 4.5
>>> ma_coefs[2] = 1.25
>>> # calculate amplitude of driving white noise
>>> sigsqr = sigmay ** 2 / cm.carma_variance(1.0, ar_roots, ma_coefs=ma_coefs)
Generate the time series:
>>> ny = 270
>>> time = np.empty(ny)
>>> dt = np.random.uniform(1.0, 3.0, ny)
>>> time[:90] = np.cumsum(dt[:90])
>>> time[90:2*90] = 180 + time[90-1] + np.cumsum(dt[90:2*90])
>>> time[2*90:] = 180 + time[2*90-1] + np.cumsum(dt[2*90:])
>>> y0 = mu + cm.carma_process(time, sigsqr, ar_roots, ma_coefs=ma_coefs)
>>> ysig = np.ones(ny) * y0.std() / 5.0 # standard deviation in measurement noise
>>> y = y0 + ysig * np.random.standard_normal(ny) # add measurement noise
Now lets choose the order of the CARMA model to use. carma_pack does this by finding the maximum-likelihood estimate of the carma models on a grid of (p,q), and then choosing (p,q) to minimize the AICc. Because the likelihood space is multimodal, carma_pack launches 100 optimizers with random starts, and picks the best one. Optimization is done using scipy.optimize, but appears to be unstable as it is not uncommon to get NaN at the MLE.
>>> carma_model = cm.CarmaModel(time, y, ysig) # create new CARMA process model
>>> pmax = 7 # only search over p < 7, q < p
>>> # njobs = -1 will use all the processes through the multiprocessing module
>>> MLE, pqlist, AICc_list = carma_model.choose_order(pmax, njobs=-1)
The chosen p and q values are saved as data members of the carma_model object. Finally, lets sample the CARMA process parameters using MCMC and the values of (p,q) chosen by the choose_order method.
>>> carma_sample = carma_model.run_mcmc(50000)
This will run the MCMC sampler for 75000 iterations, with the first 25000 of those discarded as burn-in. You can specify the number of parallel chains to use in the parallel tempering algorithm, as well as the number of iterations used by burn-in. In this example we just use the default.
Now let's examine the results. We can see which parameters we have sampled as
>>> print carma_sample.parameters
We can grab the samples for a parameter doing
>>> trace_qpo_cent = carma_sample.get_samples('psd_centroid') # get samples for the Lorentzians centroids
We can plot a useful summary of the MCMC samples for each parameter:
>>> carma_sample.plot_parameter('var') # plots for the CARMA model variance
We can also plot 1-d and 2-d posteriors:
>>> carma_sample.plot_1dpdf('ma_coefs') # histograms of MA coefficients
>>> # plot distribution of centroid of 1st Lorentzian vs width of 2nd Lorentzian
>>> carma_sample.plot_2dkde('psd_centroid', 'psd_width', 0, 1)
The pointwise 95% credibility region for the power spectrum under the CARMA model is
>>> psd_lo, psd_hi, psd_mid, freq = carma_sample.plot_power_spectrum(percentile=95.0, nsamples=5000) # only use 5000 MCMC samples for speed
Finally, we can assess the quality of the fit through
>>> carma_sample.assess_fit()
carma_pack depends on the following python modules:
numpy(for core functionality)scipy(for core functionality)matplotlib(for generating plots)acor(for calculating the autocorrelation time scale of MCMC samples)
In addition, it is necessary to have the Boost C++ libraries (for linking python and C++) and the Armadillo C++ linear algebra library installed. If you have multiple python versions installed on your system, make sure that the BoostPython links to the correct python library. Otherwise, python will crash when you import carmcmc. Note that carma_pack only works with Boost version 1.59 and earlier; it does not work with Boost version 1.60 and onwards.
Full install directions are given in the Linux_install.txt and MacOSX_install.txt files. Be forewarned that on Mac OS X getting BoostPython to build against any python other than the system python can be a real headache. You may have to replace some of the python libraries in /usr/lib with symbolic links to the actual python libraries used.
We have supplied an ipython notebook under the examples/ folder that gives a guided tour of carma_pack. Also, the script carma_pack/src/paper/carma_paper.py generates the plots
from the paper and provides additional examples of carma_pack's functionality.