-
Notifications
You must be signed in to change notification settings - Fork 2.8k
/
arima_model.py
1687 lines (1476 loc) · 64 KB
/
arima_model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from __future__ import absolute_import
# for 2to3 with extensions
from datetime import datetime
import numpy as np
from scipy import optimize
from scipy.stats import t, norm
from scipy.signal import lfilter
from numpy import (dot, identity, kron, log, zeros, pi, exp, eye, abs, empty,
zeros_like)
from numpy.linalg import inv, pinv
from statsmodels.tools.decorators import (cache_readonly,
cache_writable, resettable_cache)
import statsmodels.base.model as base
import statsmodels.tsa.base.tsa_model as tsbase
import statsmodels.base.wrapper as wrap
from statsmodels.regression.linear_model import yule_walker, GLS
from statsmodels.tsa.tsatools import (lagmat, add_trend,
_ar_transparams, _ar_invtransparams, _ma_transparams,
_ma_invtransparams)
from statsmodels.tsa.vector_ar import util
from statsmodels.tsa.ar_model import AR
from statsmodels.tsa.arima_process import arma2ma
from statsmodels.tools.numdiff import (approx_fprime, approx_fprime_cs,
approx_hess_cs)
from statsmodels.tsa.base.datetools import _index_date
from statsmodels.tsa.kalmanf import KalmanFilter
from .kalmanf import kalman_loglike
_armax_notes = """
Notes
-----
If exogenous variables are given, then the model that is fit is
.. math::
\\phi(L)(y_t - X_t\\beta) = \\theta(L)\epsilon_t
where :math:`\\phi` and :math:`\\theta` are polynomials in the lag
operator, :math:`L`. This is the regression model with ARMA errors,
or ARMAX model. This specification is used, whether or not the model
is fit using conditional sum of square or maximum-likelihood, using
the `method` argument in :meth:`statsmodels.tsa.arima_model.%(Model)s.fit`. Therefore, for now, `css`
and `mle` refer to estimation methods only. This may change for the
case of the `css` model in future versions.
"""
_arma_params = """\
endog : array-like
The endogenous variable.
order : iterable
The (p,q) order of the model for the number of AR parameters,
differences, and MA parameters to use. Though optional, the order
keyword in fit is deprecated and it is recommended to give order here.
exog : array-like, optional
An optional arry of exogenous variables. This should *not* include a
constant or trend. You can specify this in the `fit` method."""
_arma_model = "Autoregressive Moving Average ARMA(p,q) Model"
_arima_model = "Autoregressive Integrated Moving Average ARIMA(p,d,q) Model"
_arima_params = """\
endog : array-like
The endogenous variable.
order : iterable
The (p,d,q) order of the model for the number of AR parameters,
differences, and MA parameters to use.
exog : array-like, optional
An optional arry of exogenous variables. This should *not* include a
constant or trend. You can specify this in the `fit` method."""
_predict_notes = """
Notes
-----
Use the results predict method instead.
"""
_results_notes = """
Notes
-----
It is recommended to use dates with the time-series models, as the
below will probably make clear. However, if ARIMA is used without
dates and/or `start` and `end` are given as indices, then these
indices are in terms of the *original*, undifferenced series. Ie.,
given some undifferenced observations::
1970Q1, 1
1970Q2, 1.5
1970Q3, 1.25
1970Q4, 2.25
1971Q1, 1.2
1971Q2, 4.1
1970Q1 is observation 0 in the original series. However, if we fit an
ARIMA(p,1,q) model then we lose this first observation through
differencing. Therefore, the first observation we can forecast (if
using exact MLE) is index 1. In the differenced series this is index
0, but we refer to it as 1 from the original series.
"""
_predict = """
%(Model)s model in-sample and out-of-sample prediction
Parameters
----------
%(params)s
start : int, str, or datetime
Zero-indexed observation number at which to start forecasting, ie.,
the first forecast is start. Can also be a date string to
parse or a datetime type.
end : int, str, or datetime
Zero-indexed observation number at which to end forecasting, ie.,
the first forecast is start. Can also be a date string to
parse or a datetime type. However, if the dates index does not
have a fixed frequency, end must be an integer index if you
want out of sample prediction.
exog : array-like, optional
If the model is an ARMAX and out-of-sample forecasting is
requested, exog must be given. Note that you'll need to pass
`k_ar` additional lags for any exogenous variables. E.g., if you
fit an ARMAX(2, q) model and want to predict 5 steps, you need 7
observations to do this.
dynamic : bool, optional
The `dynamic` keyword affects in-sample prediction. If dynamic
is False, then the in-sample lagged values are used for
prediction. If `dynamic` is True, then in-sample forecasts are
used in place of lagged dependent variables. The first forecasted
value is `start`.
%(extra_params)s
Returns
-------
predict : array
The predicted values.
%(extra_section)s
"""
_arma_predict = _predict % {"Model" : "ARMA",
"params" : """
params : array-like
The fitted parameters of the model.""",
"extra_params" : "",
"extra_section" : _predict_notes}
_arma_results_predict = _predict % {"Model" : "ARMA", "params" : "",
"extra_params" : "", "extra_section" : _results_notes}
_arima_predict = _predict % {"Model" : "ARIMA",
"params" : """
params : array-like
The fitted parameters of the model.""",
"extra_params" :
"""
typ : str {'linear', 'levels'}
- 'linear' : Linear prediction in terms of the differenced
endogenous variables.
- 'levels' : Predict the levels of the original endogenous
variables.""",
"extra_section" : _predict_notes}
_arima_results_predict = _predict % {"Model" : "ARIMA",
"params" : "",
"extra_params" : """typ : str {'linear', 'levels'}
- 'linear' : Linear prediction in terms of the differenced
endogenous variables.
- 'levels' : Predict the levels of the original endogenous
variables.
""",
"extra_section" : _results_notes}
def _check_arima_start(start, k_ar, k_diff, method, dynamic):
if start < 0:
raise ValueError("The start index %d of the original series "
"has been differenced away" % start)
elif (dynamic or 'mle' not in method) and start < k_ar:
raise ValueError("Start must be >= k_ar for conditional MLE "
"or dynamic forecast. Got %d" % start)
def _get_predict_out_of_sample(endog, p, q, k_trend, k_exog, start, errors,
trendparam, exparams, arparams, maparams, steps,
method, exog=None):
"""
Returns endog, resid, mu of appropriate length for out of sample
prediction.
"""
if q:
resid = np.zeros(q)
if start and 'mle' in method or (start == p and not start == 0):
resid[:q] = errors[start-q:start]
elif start:
resid[:q] = errors[start-q-p:start-p]
else:
resid[:q] = errors[-q:]
else:
resid = None
y = endog
if k_trend == 1:
# use expectation not constant
if k_exog > 0:
#TODO: technically should only hold for MLE not
# conditional model. See #274.
# ensure 2-d for conformability
if np.ndim(exog) == 1 and k_exog == 1:
# have a 1d series of observations -> 2d
exog = exog[:, None]
elif np.ndim(exog) == 1:
# should have a 1d row of exog -> 2d
if len(exog) != k_exog:
raise ValueError("1d exog given and len(exog) != k_exog")
exog = exog[None, :]
X = lagmat(np.dot(exog, exparams), p, original='in', trim='both')
mu = trendparam * (1 - arparams.sum())
# arparams were reversed in unpack for ease later
mu = mu + (np.r_[1, -arparams[::-1]]*X).sum(1)[:,None]
else:
mu = trendparam * (1 - arparams.sum())
mu = np.array([mu]*steps)
else:
mu = np.zeros(steps)
endog = np.zeros(p + steps - 1)
if p and start:
endog[:p] = y[start-p:start]
elif p:
endog[:p] = y[-p:]
return endog, resid, mu
def _arma_predict_out_of_sample(params, steps, errors, p, q, k_trend, k_exog,
endog, exog=None, start=0, method='mle'):
(trendparam, exparams,
arparams, maparams) = _unpack_params(params, (p,q), k_trend,
k_exog, reverse=True)
endog, resid, mu = _get_predict_out_of_sample(endog, p, q, k_trend, k_exog,
start, errors, trendparam,
exparams, arparams,
maparams, steps, method,
exog)
forecast = np.zeros(steps)
if steps == 1:
if q:
return mu[0] + np.dot(arparams, endog[:p]) + np.dot(maparams,
resid[:q])
else:
return mu[0] + np.dot(arparams, endog[:p])
if q:
i = 0 # if q == 1
else:
i = -1
for i in range(min(q,steps-1)):
fcast = mu[i] + np.dot(arparams,endog[i:i+p]) + \
np.dot(maparams[:q-i],resid[i:i+q])
forecast[i] = fcast
endog[i+p] = fcast
for i in range(i+1,steps-1):
fcast = mu[i] + np.dot(arparams,endog[i:i+p])
forecast[i] = fcast
endog[i+p] = fcast
#need to do one more without updating endog
forecast[-1] = mu[-1] + np.dot(arparams,endog[steps-1:])
return forecast
def _arma_predict_in_sample(start, end, endog, resid, k_ar,
method):
"""
Pre- and in-sample fitting for ARMA.
"""
if 'mle' in method:
fittedvalues = endog - resid #get them all then trim
elif k_ar > 0:
fittedvalues = endog[k_ar:] - resid
fv_start = start
if 'mle' not in method:
fv_start -= k_ar # start is in terms of endog index
predictedvalues = np.zeros(end + 1 - fv_start)
fv_end = min(len(fittedvalues), end + 1)
return fittedvalues[fv_start:fv_end]
def _validate(start, k_ar, k_diff, dates, method):
if isinstance(start, (basestring, datetime)):
start_date = start
start = _index_date(start, dates)
start -= k_diff
if 'mle' not in method and start < k_ar - k_diff:
raise ValueError("Start must be >= k_ar for conditional "
"MLE or dynamic forecast. Got %s" % start)
return start
def _unpack_params(params, order, k_trend, k_exog, reverse=False):
p, q = order
k = k_trend + k_exog
maparams = params[k+p:]
arparams = params[k:k+p]
trend = params[:k_trend]
exparams = params[k_trend:k]
if reverse:
return trend, exparams, arparams[::-1], maparams[::-1]
return trend, exparams, arparams, maparams
def _unpack_order(order):
k_ar, k_ma, k = order
k_lags = max(k_ar, k_ma+1)
return k_ar, k_ma, order, k_lags
def _make_arma_names(data, k_trend, order, exog_names):
k_ar, k_ma = order
exog_names = exog_names or []
ar_lag_names = util.make_lag_names([data.ynames], k_ar, 0)
ar_lag_names = [''.join(('ar.', i))
for i in ar_lag_names]
ma_lag_names = util.make_lag_names([data.ynames], k_ma, 0)
ma_lag_names = [''.join(('ma.', i)) for i in ma_lag_names]
trend_name = util.make_lag_names('', 0, k_trend)
exog_names = trend_name + exog_names + ar_lag_names + ma_lag_names
return exog_names
def _make_arma_exog(endog, exog, trend):
k_trend = 1 # overwritten if no constant
if exog is None and trend == 'c': # constant only
exog = np.ones((len(endog),1))
elif exog is not None and trend == 'c': # constant plus exogenous
exog = add_trend(exog, trend='c', prepend=True)
elif exog is not None and trend == 'nc':
# make sure it's not holding constant from last run
if exog.var() == 0:
exog = None
k_trend = 0
if trend == 'nc':
k_trend = 0
return k_trend, exog
class ARMA(tsbase.TimeSeriesModel):
__doc__ = tsbase._tsa_doc % {"model" : _arma_model,
"params" : _arma_params, "extra_params" : "",
"extra_sections" : _armax_notes % {"Model" : "ARMA"}}
def __init__(self, endog, order=None, exog=None, dates=None, freq=None,
missing='none'):
super(ARMA, self).__init__(endog, exog, dates, freq)
exog = self.data.exog # get it after it's gone through processing
if order is None:
import warnings
warnings.warn("In the next release order will not be optional "
"in the model constructor.", FutureWarning)
else:
self.k_ar = k_ar = order[0]
self.k_ma = k_ma = order[1]
self.k_lags = k_lags = max(k_ar,k_ma+1)
if exog is not None:
if exog.ndim == 1:
exog = exog[:,None]
k_exog = exog.shape[1] # number of exog. variables excl. const
else:
k_exog = 0
self.k_exog = k_exog
def _fit_start_params_hr(self, order):
"""
Get starting parameters for fit.
Parameters
----------
order : iterable
(p,q,k) - AR lags, MA lags, and number of exogenous variables
including the constant.
Returns
-------
start_params : array
A first guess at the starting parameters.
Notes
-----
If necessary, fits an AR process with the laglength selected according
to best BIC. Obtain the residuals. Then fit an ARMA(p,q) model via
OLS using these residuals for a first approximation. Uses a separate
OLS regression to find the coefficients of exogenous variables.
References
----------
Hannan, E.J. and Rissanen, J. 1982. "Recursive estimation of mixed
autoregressive-moving average order." `Biometrika`. 69.1.
"""
p,q,k = order
start_params = zeros((p+q+k))
endog = self.endog.copy() # copy because overwritten
exog = self.exog
if k != 0:
ols_params = GLS(endog, exog).fit().params
start_params[:k] = ols_params
endog -= np.dot(exog, ols_params).squeeze()
if q != 0:
if p != 0:
armod = AR(endog).fit(ic='bic', trend='nc')
arcoefs_tmp = armod.params
p_tmp = armod.k_ar
# it's possible in small samples that optimal lag-order
# doesn't leave enough obs. No consistent way to fix.
if p_tmp + q >= len(endog):
raise ValueError("Proper starting parameters cannot"
" be found for this order with this number "
"of observations. Use the start_params "
"argument.")
resid = endog[p_tmp:] - np.dot(lagmat(endog, p_tmp,
trim='both'), arcoefs_tmp)
if p < p_tmp + q:
endog_start = p_tmp + q - p
resid_start = 0
else:
endog_start = 0
resid_start = p - p_tmp - q
lag_endog = lagmat(endog, p, 'both')[endog_start:]
lag_resid = lagmat(resid, q, 'both')[resid_start:]
# stack ar lags and resids
X = np.column_stack((lag_endog, lag_resid))
coefs = GLS(endog[max(p_tmp+q,p):], X).fit().params
start_params[k:k+p+q] = coefs
else:
start_params[k+p:k+p+q] = yule_walker(endog, order=q)[0]
if q == 0 and p != 0:
arcoefs = yule_walker(endog, order=p)[0]
start_params[k:k+p] = arcoefs
# check AR coefficients
if p and not np.all(np.abs(np.roots(np.r_[1,
-start_params[k:k+p]])) < 1):
raise ValueError("The computed initial AR coefficients are not "
"stationary\nYou should induce stationarity, "
"choose a different model order, or you can\n"
"pass your own start_params.")
# check MA coefficients
elif q and not np.all(np.abs(np.roots(np.r_[1,
start_params[k+p:]])) < 1):
raise ValueError("The computed initial MA coefficients are not "
"invertible\nYou should induce invertibility, "
"choose a different model order, or you can\n"
"pass your own start_params.")
# check MA coefficients
return start_params
def _fit_start_params(self, order, method):
if method != 'css-mle': # use Hannan-Rissanen to get start params
start_params = self._fit_start_params_hr(order)
else: # use CSS to get start params
func = lambda params: -self.loglike_css(params)
#start_params = [.1]*(k_ar+k_ma+k_exog) # different one for k?
start_params = self._fit_start_params_hr(order)
if self.transparams:
start_params = self._invtransparams(start_params)
bounds = [(None,)*2]*sum(order)
mlefit = optimize.fmin_l_bfgs_b(func, start_params,
approx_grad=True, m=12, pgtol=1e-7, factr=1e3,
bounds = bounds, iprint=-1)
start_params = self._transparams(mlefit[0])
return start_params
def score(self, params):
"""
Compute the score function at params.
Notes
-----
This is a numerical approximation.
"""
loglike = self.loglike
#if self.transparams:
# params = self._invtransparams(params)
#return approx_fprime(params, loglike, epsilon=1e-5)
return approx_fprime_cs(params, loglike)
def hessian(self, params):
"""
Compute the Hessian at params,
Notes
-----
This is a numerical approximation.
"""
loglike = self.loglike
#if self.transparams:
# params = self._invtransparams(params)
return approx_hess_cs(params, loglike)
def _transparams(self, params):
"""
Transforms params to induce stationarity/invertability.
Reference
---------
Jones(1980)
"""
k_ar, k_ma = self.k_ar, self.k_ma
k = self.k_exog + self.k_trend
newparams = np.zeros_like(params)
# just copy exogenous parameters
if k != 0:
newparams[:k] = params[:k]
# AR Coeffs
if k_ar != 0:
newparams[k:k+k_ar] = _ar_transparams(params[k:k+k_ar].copy())
# MA Coeffs
if k_ma != 0:
newparams[k+k_ar:] = _ma_transparams(params[k+k_ar:].copy())
return newparams
def _invtransparams(self, start_params):
"""
Inverse of the Jones reparameterization
"""
k_ar, k_ma = self.k_ar, self.k_ma
k = self.k_exog + self.k_trend
newparams = start_params.copy()
arcoefs = newparams[k:k+k_ar]
macoefs = newparams[k+k_ar:]
# AR coeffs
if k_ar != 0:
newparams[k:k+k_ar] = _ar_invtransparams(arcoefs)
# MA coeffs
if k_ma != 0:
newparams[k+k_ar:k+k_ar+k_ma] = _ma_invtransparams(macoefs)
return newparams
def _get_predict_start(self, start, dynamic):
# do some defaults
method = getattr(self, 'method', 'mle')
k_ar = getattr(self, 'k_ar', 0)
k_diff = getattr(self, 'k_diff', 0)
if start is None:
if 'mle' in method and not dynamic:
start = 0
else:
start = k_ar
self._set_predict_start_date(start) # else it's done in super
elif isinstance(start, int):
start = super(ARMA, self)._get_predict_start(start)
else: # should be on a date
#elif 'mle' not in method or dynamic: # should be on a date
start = _validate(start, k_ar, k_diff, self.data.dates,
method)
start = super(ARMA, self)._get_predict_start(start)
_check_arima_start(start, k_ar, k_diff, method, dynamic)
return start
def _get_predict_end(self, end, dynamic=False):
# pass through so predict works for ARIMA and ARMA
return super(ARMA, self)._get_predict_end(end)
def geterrors(self, params):
"""
Get the errors of the ARMA process.
Parameters
----------
params : array-like
The fitted ARMA parameters
order : array-like
3 item iterable, with the number of AR, MA, and exogenous
parameters, including the trend
"""
#start = self._get_predict_start(start) # will be an index of a date
#end, out_of_sample = self._get_predict_end(end)
params = np.asarray(params)
k_ar, k_ma = self.k_ar, self.k_ma
k = self.k_exog + self.k_trend
method = getattr(self, 'method', 'mle')
if 'mle' in method: # use KalmanFilter to get errors
(y, k, nobs, k_ar, k_ma, k_lags, newparams, Z_mat, m, R_mat,
T_mat, paramsdtype) = KalmanFilter._init_kalman_state(params, self)
errors = KalmanFilter.geterrors(y,k,k_ar,k_ma, k_lags, nobs,
Z_mat, m, R_mat, T_mat, paramsdtype)
if isinstance(errors, tuple):
errors = errors[0] # non-cython version returns a tuple
else: # use scipy.signal.lfilter
y = self.endog.copy()
k = self.k_exog + self.k_trend
if k > 0:
y -= dot(self.exog, params[:k])
k_ar = self.k_ar
k_ma = self.k_ma
(trendparams, exparams,
arparams, maparams) = _unpack_params(params, (k_ar, k_ma),
self.k_trend, self.k_exog,
reverse=False)
b,a = np.r_[1,-arparams], np.r_[1,maparams]
zi = zeros((max(k_ar, k_ma)))
for i in range(k_ar):
zi[i] = sum(-b[:i+1][::-1]*y[:i+1])
e = lfilter(b,a,y,zi=zi)
errors = e[0][k_ar:]
return errors.squeeze()
def predict(self, params, start=None, end=None, exog=None, dynamic=False):
method = getattr(self, 'method', 'mle') # don't assume fit
#params = np.asarray(params)
# will return an index of a date
start = self._get_predict_start(start, dynamic)
end, out_of_sample = self._get_predict_end(end, dynamic)
if out_of_sample and (exog is None and self.k_exog > 0):
raise ValueError("You must provide exog for ARMAX")
endog = self.endog
resid = self.geterrors(params)
k_ar = self.k_ar
if out_of_sample != 0 and self.k_exog > 0:
if self.k_exog == 1 and exog.ndim == 1:
exog = exog[:,None]
# we need the last k_ar exog for the lag-polynomial
if self.k_exog > 0:
# need the last k_ar exog for the lag-polynomial
exog = np.vstack((self.exog[-k_ar:, self.k_trend:], exog))
if dynamic:
#TODO: now that predict does dynamic in-sample it should
# also return error estimates and confidence intervals
# but how? len(endog) is not tot_obs
out_of_sample += end - start + 1
return _arma_predict_out_of_sample(params, out_of_sample, resid,
k_ar, self.k_ma, self.k_trend, self.k_exog, endog, exog,
start, method)
predictedvalues = _arma_predict_in_sample(start, end, endog, resid,
k_ar, method)
if out_of_sample:
forecastvalues = _arma_predict_out_of_sample(params, out_of_sample,
resid, k_ar, self.k_ma, self.k_trend,
self.k_exog, endog, exog,
method=method)
predictedvalues = np.r_[predictedvalues, forecastvalues]
return predictedvalues
predict.__doc__ = _arma_predict
def loglike(self, params):
"""
Compute the log-likelihood for ARMA(p,q) model
Notes
-----
Likelihood used depends on the method set in fit
"""
method = self.method
if method in ['mle', 'css-mle']:
return self.loglike_kalman(params)
elif method == 'css':
return self.loglike_css(params)
else:
raise ValueError("Method %s not understood" % method)
def loglike_kalman(self, params):
"""
Compute exact loglikelihood for ARMA(p,q) model using the Kalman Filter.
"""
return KalmanFilter.loglike(params, self)
def loglike_css(self, params):
"""
Conditional Sum of Squares likelihood function.
"""
k_ar = self.k_ar
k_ma = self.k_ma
k = self.k_exog + self.k_trend
y = self.endog.copy().astype(params.dtype)
nobs = self.nobs
# how to handle if empty?
if self.transparams:
newparams = self._transparams(params)
else:
newparams = params
if k > 0:
y -= dot(self.exog, newparams[:k])
# the order of p determines how many zeros errors to set for lfilter
b,a = np.r_[1,-newparams[k:k+k_ar]], np.r_[1,newparams[k+k_ar:]]
zi = np.zeros((max(k_ar,k_ma)), dtype=params.dtype)
for i in range(k_ar):
zi[i] = sum(-b[:i+1][::-1] * y[:i+1])
errors = lfilter(b,a, y, zi=zi)[0][k_ar:]
ssr = np.dot(errors,errors)
sigma2 = ssr/nobs
self.sigma2 = sigma2
llf = -nobs/2.*(log(2*pi) + log(sigma2)) - ssr/(2*sigma2)
return llf
def fit(self, order=None, start_params=None, trend='c', method = "css-mle",
transparams=True, solver=None, maxiter=35, full_output=1,
disp=5, callback=None, **kwargs):
"""
Fits ARMA(p,q) model using exact maximum likelihood via Kalman filter.
Parameters
----------
start_params : array-like, optional
Starting parameters for ARMA(p,q). If None, the default is given
by ARMA._fit_start_params. See there for more information.
transparams : bool, optional
Whehter or not to transform the parameters to ensure stationarity.
Uses the transformation suggested in Jones (1980). If False,
no checking for stationarity or invertibility is done.
method : str {'css-mle','mle','css'}
This is the loglikelihood to maximize. If "css-mle", the
conditional sum of squares likelihood is maximized and its values
are used as starting values for the computation of the exact
likelihood via the Kalman filter. If "mle", the exact likelihood
is maximized via the Kalman Filter. If "css" the conditional sum
of squares likelihood is maximized. All three methods use
`start_params` as starting parameters. See above for more
information.
trend : str {'c','nc'}
Whehter to include a constant or not. 'c' includes constant,
'nc' no constant.
solver : str or None, optional
Solver to be used. The default is 'l_bfgs' (limited memory
Broyden-Fletcher-Goldfarb-Shanno). Other choices are 'bfgs',
'newton' (Newton-Raphson), 'nm' (Nelder-Mead), 'cg' -
(conjugate gradient), 'ncg' (non-conjugate gradient), and
'powell'. By default, the limited memory BFGS uses m=12 to
approximate the Hessian, projected gradient tolerance of 1e-8 and
factr = 1e2. You can change these by using kwargs.
maxiter : int, optional
The maximum number of function evaluations. Default is 35.
tol : float
The convergence tolerance. Default is 1e-08.
full_output : bool, optional
If True, all output from solver will be available in
the Results object's mle_retvals attribute. Output is dependent
on the solver. See Notes for more information.
disp : bool, optional
If True, convergence information is printed. For the default
l_bfgs_b solver, disp controls the frequency of the output during
the iterations. disp < 0 means no output in this case.
callback : function, optional
Called after each iteration as callback(xk) where xk is the current
parameter vector.
kwargs
See Notes for keyword arguments that can be passed to fit.
Returns
-------
statsmodels.tsa.arima_model.ARMAResults class
See also
--------
statsmodels.base.model.LikelihoodModel.fit : for more information
on using the solvers.
ARMAResults : results class returned by fit
Notes
------
If fit by 'mle', it is assumed for the Kalman Filter that the initial
unkown state is zero, and that the inital variance is
P = dot(inv(identity(m**2)-kron(T,T)),dot(R,R.T).ravel('F')).reshape(r,
r, order = 'F')
"""
if order is not None:
import warnings
warnings.warn("The order argument to fit is deprecated. "
"Please use the model constructor argument order. "
"This will overwrite any order given in the model "
"constructor.", FutureWarning)
# get model order and constants
self.k_ar = k_ar = int(order[0])
self.k_ma = k_ma = int(order[1])
self.k_lags = max(k_ar,k_ma+1)
else:
try:
assert hasattr(self, "k_ar")
assert hasattr(self, "k_ma")
except:
raise ValueError("Please give order to the model constructor "
"before calling fit.")
k_ar = self.k_ar
k_ma = self.k_ma
# enforce invertibility
self.transparams = transparams
self.method = method.lower()
endog, exog = self.endog, self.exog
k_exog = self.k_exog
self.nobs = len(endog) # this is overwritten if method is 'css'
# (re)set trend and handle exogenous variables
# always pass original exog
k_trend, exog = _make_arma_exog(endog, self.exog, trend)
self.k_trend = k_trend
self.exog = exog # overwrites original exog from __init__
# (re)set names for this model
self.exog_names = _make_arma_names(self.data, k_trend, (k_ar, k_ma),
self.exog_names)
k = k_trend + k_exog
# choose objective function
method = method.lower()
# adjust nobs for css
if method == 'css':
self.nobs = len(self.endog) - k_ar
loglike = lambda params: -self.loglike(params)
if start_params is not None:
start_params = np.asarray(start_params)
else: # estimate starting parameters
start_params = self._fit_start_params((k_ar,k_ma,k), method)
if transparams: # transform initial parameters to ensure invertibility
start_params = self._invtransparams(start_params)
if solver is None: # use default limited memory bfgs
bounds = [(None,)*2]*(k_ar+k_ma+k)
pgtol = kwargs.get('pgtol', 1e-8)
factr = kwargs.get('factr', 1e2)
m = kwargs.get('m', 12)
mlefit = optimize.fmin_l_bfgs_b(loglike, start_params,
approx_grad=True, m=m, pgtol=pgtol, factr=factr,
bounds=bounds, iprint=disp)
self.mlefit = mlefit
params = mlefit[0]
else: # call the solver from LikelihoodModel
mlefit = super(ARMA, self).fit(start_params, method=solver,
maxiter=maxiter, full_output=full_output, disp=disp,
callback = callback, **kwargs)
self.mlefit = mlefit
params = mlefit.params
if transparams: # transform parameters back
params = self._transparams(params)
self.transparams = False # set to false so methods don't expect transf.
normalized_cov_params = None #TODO: fix this
armafit = ARMAResults(self, params, normalized_cov_params)
return ARMAResultsWrapper(armafit)
#NOTE: the length of endog changes when we give a difference to fit
#so model methods are not the same on unfit models as fit ones
#starting to think that order of model should be put in instantiation...
class ARIMA(ARMA):
__doc__ = tsbase._tsa_doc % {"model" : _arima_model,
"params" : _arima_params, "extra_params" : "",
"extra_sections" : _armax_notes % {"Model" : "ARIMA"}}
def __new__(cls, endog, order, exog=None, dates=None, freq=None,
missing='none'):
p, d, q = order
if d == 0: # then we just use an ARMA model
return ARMA(endog, (p,q), exog, dates, freq, missing)
else:
mod = super(ARIMA, cls).__new__(cls)
mod.__init__(endog, order, exog, dates, freq, missing)
return mod
def __init__(self, endog, order, exog=None, dates=None, freq=None,
missing='none'):
p,d,q = order
super(ARIMA, self).__init__(endog, (p,q), exog, dates, freq, missing)
self.k_diff = d
self.endog = np.diff(self.endog, n=d)
if exog is not None:
self.exog = self.exog[d:]
self.data.ynames = 'D.' + self.endog_names
# what about exog, should we difference it automatically before
# super call?
def _get_predict_start(self, start, dynamic):
"""
"""
#TODO: remove all these getattr and move order specification to
# class constructor
k_diff = getattr(self, 'k_diff', 0)
method = getattr(self, 'method', 'mle')
k_ar = getattr(self, 'k_ar', 0)
if start is None:
if 'mle' in method and not dynamic:
start = 0
else:
start = k_ar
elif isinstance(start, int):
start -= k_diff
try: # catch when given an integer outside of dates index
start = super(ARIMA, self)._get_predict_start(start,
dynamic)
except IndexError, err:
raise ValueError("start must be in series. "
"got %d" % (start + k_diff))
else: # received a date
start = _validate(start, k_ar, k_diff, self.data.dates,
method)
start = super(ARIMA, self)._get_predict_start(start, dynamic)
# reset date for k_diff adjustment
self._set_predict_start_date(start + k_diff)
return start
def _get_predict_end(self, end, dynamic=False):
"""
Returns last index to be forecast of the differenced array.
Handling of inclusiveness should be done in the predict function.
"""
end, out_of_sample = super(ARIMA, self)._get_predict_end(end, dynamic)
if 'mle' not in self.method and not dynamic:
end -= self.k_ar
return end - self.k_diff, out_of_sample
def fit(self, start_params=None, trend='c', method = "css-mle",
transparams=True, solver=None, maxiter=35, full_output=1,
disp=5, callback=None, **kwargs):
"""
Fits ARIMA(p,d,q) model by exact maximum likelihood via Kalman filter.
Parameters
----------
start_params : array-like, optional
Starting parameters for ARMA(p,q). If None, the default is given
by ARMA._fit_start_params. See there for more information.
transparams : bool, optional
Whehter or not to transform the parameters to ensure stationarity.
Uses the transformation suggested in Jones (1980). If False,
no checking for stationarity or invertibility is done.
method : str {'css-mle','mle','css'}
This is the loglikelihood to maximize. If "css-mle", the
conditional sum of squares likelihood is maximized and its values
are used as starting values for the computation of the exact
likelihood via the Kalman filter. If "mle", the exact likelihood
is maximized via the Kalman Filter. If "css" the conditional sum
of squares likelihood is maximized. All three methods use
`start_params` as starting parameters. See above for more
information.
trend : str {'c','nc'}
Whether to include a constant or not. 'c' includes constant,
'nc' no constant.
solver : str or None, optional
Solver to be used. The default is 'l_bfgs' (limited memory
Broyden-Fletcher-Goldfarb-Shanno). Other choices are 'bfgs',
'newton' (Newton-Raphson), 'nm' (Nelder-Mead), 'cg' -
(conjugate gradient), 'ncg' (non-conjugate gradient), and
'powell'. By default, the limited memory BFGS uses m=12 to
approximate the Hessian, projected gradient tolerance of 1e-8 and
factr = 1e2. You can change these by using kwargs.
maxiter : int, optional
The maximum number of function evaluations. Default is 35.
tol : float
The convergence tolerance. Default is 1e-08.
full_output : bool, optional
If True, all output from solver will be available in
the Results object's mle_retvals attribute. Output is dependent
on the solver. See Notes for more information.
disp : bool, optional
If True, convergence information is printed. For the default
l_bfgs_b solver, disp controls the frequency of the output during
the iterations. disp < 0 means no output in this case.
callback : function, optional
Called after each iteration as callback(xk) where xk is the current
parameter vector.
kwargs
See Notes for keyword arguments that can be passed to fit.
Returns
-------
`statsmodels.tsa.arima.ARIMAResults` class