def test_compare_arma():
#this is a preliminary test to compare arma_kf, arma_cond_ls and arma_cond_mle
#the results returned by the fit methods are incomplete
#for now without random.seed
#np.random.seed(9876565)
x = fa.ArmaFft([1, -0.5], [1., 0.4], 40).generate_sample(size=200,
burnin=1000)
# this used kalman filter through descriptive
# d = ARMA(x)
# d.fit((1,1), trend='nc')
# dres = d.res
modkf = ARMA(x)
##rkf = mkf.fit((1,1))
##rkf.params
reskf = modkf.fit((1,1), trend='nc', disp=-1)
dres = reskf
modc = Arma(x)
resls = modc.fit(order=(1,1))
rescm = modc.fit_mle(order=(1,1), start_params=[0.4,0.4, 1.], disp=0)
#decimal 1 corresponds to threshold of 5% difference
#still different sign corrcted
#assert_almost_equal(np.abs(resls[0] / d.params), np.ones(d.params.shape), decimal=1)
assert_almost_equal(resls[0] / dres.params, np.ones(dres.params.shape),
decimal=1)
#rescm also contains variance estimate as last element of params
#assert_almost_equal(np.abs(rescm.params[:-1] / d.params), np.ones(d.params.shape), decimal=1)
assert_almost_equal(rescm.params[:-1] / dres.params, np.ones(dres.params.shape), decimal=1)
def mcarma22(niter=10, nsample=1000, ar=None, ma=None, sig=0.5):
'''run Monte Carlo for ARMA(2,2)
DGP parameters currently hard coded
also sample size `nsample`
was not a self contained function, used instances from outer scope
now corrected
'''
#nsample = 1000
#ar = [1.0, 0, 0]
if ar is None:
ar = [1.0, -0.55, -0.1]
#ma = [1.0, 0, 0]
if ma is None:
ma = [1.0, 0.3, 0.2]
results = []
results_bse = []
for _ in range(niter):
y2 = arma_generate_sample(ar,ma,nsample+1000, sig)[-nsample:]
y2 -= y2.mean()
arest2 = Arma(y2)
rhohat2a, cov_x2a, infodict, mesg, ier = arest2.fit((2,2))
results.append(rhohat2a)
err2a = arest2.geterrors(rhohat2a)
sige2a = np.sqrt(np.dot(err2a,err2a)/nsample)
#print 'sige2a', sige2a,
#print 'cov_x2a.shape', cov_x2a.shape
#results_bse.append(sige2a * np.sqrt(np.diag(cov_x2a)))
if not cov_x2a is None:
results_bse.append(sige2a * np.sqrt(np.diag(cov_x2a)))
else:
results_bse.append(np.nan + np.zeros_like(rhohat2a))
return np.r_[ar[1:], ma[1:]], np.array(results), np.array(results_bse)
def mcarma22(niter=10, nsample=1000, ar=None, ma=None, sig=0.5):
'''run Monte Carlo for ARMA(2,2)
DGP parameters currently hard coded
also sample size `nsample`
was not a self contained function, used instances from outer scope
now corrected
'''
#nsample = 1000
#ar = [1.0, 0, 0]
if ar is None:
ar = [1.0, -0.55, -0.1]
#ma = [1.0, 0, 0]
if ma is None:
ma = [1.0, 0.3, 0.2]
results = []
results_bse = []
for _ in range(niter):
y2 = arma_generate_sample(ar, ma, nsample + 1000, sig)[-nsample:]
y2 -= y2.mean()
arest2 = Arma(y2)
rhohat2a, cov_x2a, infodict, mesg, ier = arest2.fit((2, 2))
results.append(rhohat2a)
err2a = arest2.geterrors(rhohat2a)
sige2a = np.sqrt(np.dot(err2a, err2a) / nsample)
#print 'sige2a', sige2a,
#print 'cov_x2a.shape', cov_x2a.shape
#results_bse.append(sige2a * np.sqrt(np.diag(cov_x2a)))
if not cov_x2a is None:
results_bse.append(sige2a * np.sqrt(np.diag(cov_x2a)))
else:
results_bse.append(np.nan + np.zeros_like(rhohat2a))
return np.r_[ar[1:], ma[1:]], np.array(results), np.array(results_bse)
def test_compare_arma():
#this is a preliminary test to compare arma_kf, arma_cond_ls and arma_cond_mle
#the results returned by the fit methods are incomplete
#for now without random.seed
#np.random.seed(9876565)
x = fa.ArmaFft([1, -0.5], [1., 0.4], 40).generate_sample(size=200,
burnin=1000)
# this used kalman filter through descriptive
# d = ARMA(x)
# d.fit((1,1), trend='nc')
# dres = d.res
modkf = ARMA(x)
##rkf = mkf.fit((1,1))
##rkf.params
reskf = modkf.fit((1,1), trend='nc', disp=-1)
dres = reskf
modc = Arma(x)
resls = modc.fit(order=(1,1))
rescm = modc.fit_mle(order=(1,1), start_params=[0.4,0.4, 1.], disp=0)
#decimal 1 corresponds to threshold of 5% difference
#still different sign corrcted
#assert_almost_equal(np.abs(resls[0] / d.params), np.ones(d.params.shape), decimal=1)
assert_almost_equal(resls[0] / dres.params, np.ones(dres.params.shape),
decimal=1)
#rescm also contains variance estimate as last element of params
#assert_almost_equal(np.abs(rescm.params[:-1] / d.params), np.ones(d.params.shape), decimal=1)
assert_almost_equal(rescm.params[:-1] / dres.params, np.ones(dres.params.shape), decimal=1)
print 'time used:', t2-t1 print "Arma.fit_mle results" # have to set nar and nma manually arma1.nar = 2 arma1.nma = 2 t2=time() ret = arma1.fit_mle() t3=time() print "params, first 4, sigma, last 1 ", ret.params results += ["Arma.fit_mle ", ret.params[:4], ret.params[-1], ret.llf] print 'time used:', t3-t2 print "Arma.fit method = \"ls\"" t3=time() ret2 = arma1.fit(order=(2,0,2), method="ls") t4=time() print ret2[0] results += ["Arma.fit ls", ret2[0]] print 'time used:', t4-t3 print "Arma.fit method = \"CLS\"" t4=time() ret3 = arma1.fit(order=(2,0,2), method="None") t5=time() print ret3 results += ["Arma.fit other", ret3[0]] print 'time used:', t5-t4 for i in results: print i
ar = [1.0, -0.8]
ma = [1.0, 0.5]
y1 = arma_generate_sample(ar, ma, 1000, 0.1)
y1 -= y1.mean() #no mean correction/constant in estimation so far
arma1 = Arma(y1)
arma1.nar = 1
arma1.nma = 1
arma1res = arma1.fit_mle(order=(1, 1), method='fmin')
print arma1res.params
#Warning need new instance otherwise results carry over
arma2 = Arma(y1)
arma2.nar = 1
arma2.nma = 1
res2 = arma2.fit(method='bfgs')
print res2.params
print res2.model.hessian(res2.params)
print ndt.Hessian(arma1.loglike, stepMax=1e-2)(res2.params)
arest = tsa.arima.ARIMA(y1)
resls = arest.fit((1, 0, 1))
print resls[0]
print resls[1]
print '\nparameter estimate - comparing methods'
print '---------------------------------------'
print 'parameter of DGP ar(1), ma(1), sigma_error'
print[-0.8, 0.5, 0.1]
print 'mle with fmin'
print arma1res.params
print 'mle with bfgs'
ar = [1.0, -0.8]
ma = [1.0, 0.5]
y1 = arma_generate_sample(ar,ma,1000,0.1)
y1 -= y1.mean() #no mean correction/constant in estimation so far
arma1 = Arma(y1)
arma1.nar = 1
arma1.nma = 1
arma1res = arma1.fit_mle(order=(1,1), method='fmin')
print arma1res.params
#Warning need new instance otherwise results carry over
arma2 = Arma(y1)
arma2.nar = 1
arma2.nma = 1
res2 = arma2.fit(method='bfgs')
print res2.params
print res2.model.hessian(res2.params)
print ndt.Hessian(arma1.loglike, stepMax=1e-2)(res2.params)
arest = tsa.arima.ARIMA(y1)
resls = arest.fit((1,0,1))
print resls[0]
print resls[1]
print '\nparameter estimate - comparing methods'
print '---------------------------------------'
print 'parameter of DGP ar(1), ma(1), sigma_error'
print [-0.8, 0.5, 0.1]
print 'mle with fmin'
from numpy.testing import assert_almost_equal import matplotlib.pyplot as plt import scikits.statsmodels.sandbox.tsa.fftarma as fa from scikits.statsmodels.tsa.descriptivestats import TsaDescriptive from scikits.statsmodels.tsa.arma_mle import Arma x = fa.ArmaFft([1, -0.5], [1., 0.4], 40).generate_sample(size=200, burnin=1000) d = TsaDescriptive(x) d.plot4() #d.fit(order=(1,1)) d.fit((1,1), trend='nc') print d.res.params modc = Arma(x) resls = modc.fit(order=(1,1)) print resls[0] rescm = modc.fit_mle(order=(1,1), start_params=[-0.4,0.4, 1.]) print rescm.params #decimal 1 corresponds to threshold of 5% difference assert_almost_equal(resls[0] / d.res.params, 1, decimal=1) assert_almost_equal(rescm.params[:-1] / d.res.params, 1, decimal=1) #copied to tsa.tests plt.figure() plt.plot(x, 'b-o') plt.plot(modc.predicted(), 'r-') plt.figure() plt.plot(modc.error_estimate) #plt.show()
from numpy.testing import assert_almost_equal import matplotlib.pyplot as plt import scikits.statsmodels.sandbox.tsa.fftarma as fa from scikits.statsmodels.tsa.descriptivestats import TsaDescriptive from scikits.statsmodels.tsa.arma_mle import Arma x = fa.ArmaFft([1, -0.5], [1., 0.4], 40).generate_sample(size=200, burnin=1000) d = TsaDescriptive(x) d.plot4() #d.fit(order=(1,1)) d.fit((1,1), trend='nc') print d.res.params modc = Arma(x) resls = modc.fit(order=(1,1)) print resls[0] rescm = modc.fit_mle(order=(1,1), start_params=[-0.4,0.4, 1.]) print rescm.params #decimal 1 corresponds to threshold of 5% difference assert_almost_equal(resls[0] / d.res.params, 1, decimal=1) assert_almost_equal(rescm.params[:-1] / d.res.params, 1, decimal=1) #copied to tsa.tests plt.figure() plt.plot(x, 'b-o') plt.plot(modc.predicted(), 'r-') plt.figure() plt.plot(modc.error_estimate) #plt.show()
#need to iterate, ar1 too large ma terms too small
#fix large parameters, if hannan_rissannen are too large
start_params_mle[:-1] = (np.sign(start_params_mle[:-1])
* np.minimum(np.abs(start_params_mle[:-1]),0.75))
print 'conditional least-squares'
#print rhohat2
print 'with mle'
arest2.nar = 2
arest2.nma = 2
#
res = arest2.fit_mle(start_params=start_params_mle, method='nm') #no order in fit
print res.params
rhohat2, cov_x2a, infodict, mesg, ier = arest2.fit((2,2))
print '\nARIMA_old'
arest = ARIMA_old(y22)
rhohat1, cov_x1, infodict, mesg, ier = arest.fit((2,0,2))
print rhohat1
print np.sqrt(np.diag(cov_x1))
err1 = arest.errfn(x=y22)
print np.var(err1)
print 'bse ls, formula not checked'
print np.sqrt(np.diag(cov_x1))*err1.std()
print 'bsejac for mle'
#print arest2.bsejac
#TODO:check bsejac raises singular matrix linalg error
#in model.py line620: return np.linalg.inv(np.dot(jacv.T, jacv))
print '\nyule-walker'
