1 /* PSPP - a program for statistical analysis.
2 Copyright (C) 2006 Free Software Foundation, Inc.
4 This program is free software: you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation, either version 3 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program. If not, see <http://www.gnu.org/licenses/>. */
18 Find preliminary ARMA coefficients via the innovations algorithm.
19 Also compute the sample mean and covariance matrix for each series.
23 P. J. Brockwell and R. A. Davis. Time Series: Theory and
24 Methods. Second edition. Springer. New York. 1991. ISBN
25 0-387-97429-6. Sections 5.2, 8.3 and 8.4.
33 #include <gsl/gsl_matrix.h>
34 #include <gsl/gsl_vector.h>
35 #include <libpspp/compiler.h>
36 #include <libpspp/misc.h>
37 #include <math/coefficient.h>
38 #include <math/ts/innovations.h>
43 get_mean (const gsl_matrix *data,
44 struct innovations_estimate **est)
52 for (n = 0; n < data->size2; n++)
57 for (i = 0; i < data->size1; i++)
59 for (n = 0; n < data->size2; n++)
61 tmp = gsl_matrix_get (data, i, n);
65 d = (tmp - est[n]->mean) / est[n]->n_obs;
72 update_cov (struct innovations_estimate **est, gsl_vector_const_view x,
73 gsl_vector_const_view y, size_t lag)
79 for (j = 0; j < x.vector.size; j++)
81 xj = gsl_vector_get (&x.vector, j);
82 yj = gsl_vector_get (&y.vector, j);
89 *(est[j]->cov + lag) += xj * yj;
95 get_covariance (const gsl_matrix *data,
96 struct innovations_estimate **est, size_t max_lag)
103 assert (data != NULL);
104 assert (est != NULL);
106 for (j = 0; j < data->size2; j++)
108 for (lag = 0; lag <= max_lag; lag++)
110 *(est[j]->cov + lag) = 0.0;
114 The rows are in the outer loop because a gsl_matrix is stored in
117 for (i = 0; i < data->size1; i++)
119 for (lag = 0; lag <= max_lag && lag < data->size1 - i; lag++)
121 update_cov (est, gsl_matrix_const_row (data, i),
122 gsl_matrix_const_row (data, i + lag), lag);
125 for (j = 0; j < data->size2; j++)
127 for (lag = 0; lag <= max_lag; lag++)
129 *(est[j]->cov + lag) /= est[j]->n_obs;
137 innovations_convolve (double *x, double *y, struct innovations_estimate *est,
143 assert (x != NULL && y != NULL);
144 assert (est != NULL);
145 assert (est->scale != NULL);
147 for (k = 0; k < i; k++)
149 result += x[k] * y[k] * est->scale[i-k-1];
154 innovations_update_scale (struct innovations_estimate *est, double *theta,
161 if (i < (size_t) est->max_lag)
163 result = est->cov[0];
164 for (j = 0; j < i; j++)
167 result -= pow2 (theta[k]) * est->scale[j];
169 est->scale[i] = result;
173 init_theta (double **theta, size_t max_lag)
178 for (i = 0; i < max_lag; i++)
180 for (j = 0; j <= i; j++)
187 innovations_update_coeff (double **theta, struct innovations_estimate *est,
194 for (i = 0; i < max_lag; i++)
196 theta[i][i] = est->cov[i+1] / est->scale[0];
197 for (j = 1; j <= i; j++)
200 theta[i][k] = (est->cov[k+1] -
201 innovations_convolve (theta[i] + k + 1, theta[j - 1], est, j))
204 innovations_update_scale (est, theta[i], i + 1);
208 get_coef (const gsl_matrix *data,
209 struct innovations_estimate **est, size_t max_lag)
215 theta = xnmalloc (max_lag, sizeof (*theta));
216 for (i = 0; i < max_lag; i++)
218 theta[i] = xnmalloc (max_lag, sizeof (**(theta + i)));
221 for (n = 0; n < data->size2; n++)
223 init_theta (theta, max_lag);
224 innovations_update_scale (est[n], theta[0], 0);
225 innovations_update_coeff (theta, est[n], max_lag);
226 /* Copy the final row of coefficients into EST->COEFF.*/
227 for (i = 0; i < max_lag; i++)
230 The order of storage here means that the best predicted value
231 for the time series is computed as follows:
233 Let X[m], X[m-1],... denote the original series.
234 Let X_hat[0] denote the best predicted value of X[0],
235 X_hat[1] denote the projection of X[1] onto the subspace
236 spanned by {X[0] - X_hat[0]}. Let X_hat[m] denote the
237 projection of X[m] onto the subspace spanned by {X[m-1] - X_hat[m-1],
238 X[m-2] - X_hat[m-2],...,X[0] - X_hat[0]}.
240 Then X_hat[m] = est->coeff[m-1] * (X[m-1] - X_hat[m-1])
241 + est->coeff[m-1] * (X[m-2] - X_hat[m-2])
243 + est->coeff[m-max_lag] * (X[m - max_lag] - X_hat[m - max_lag])
245 pspp_coeff_set_estimate (est[n]->coeff[i], theta[max_lag - 1][i]);
249 for (i = 0; i < max_lag; i++)
257 innovations_struct_init (struct innovations_estimate *est,
258 const struct design_matrix *dm,
264 /* COV[0] stores the lag 0 covariance (i.e., the variance), COV[1]
265 holds the lag-1 covariance, etc.
267 est->cov = xnmalloc (lag + 1, sizeof (*est->cov));
268 est->scale = xnmalloc (lag + 1, sizeof (*est->scale));
269 est->coeff = xnmalloc (lag, sizeof (*est->coeff)); /* No intercept. */
272 The loop below is an unusual use of PSPP_COEFF_INIT(). In a
273 typical model, one column of a DESIGN_MATRIX has one
274 coefficient. But in a time-series model, one column has many
277 for (j = 0; j < lag; j++)
279 pspp_coeff_init (est->coeff + j, dm);
281 est->max_lag = (double) lag;
284 The mean is subtracted from the original data before computing the
285 coefficients. The mean is NOT added back, so if you want to predict
286 a new value, you must add the mean to X_hat[m] to get the correct
290 subtract_mean (gsl_matrix *m, struct innovations_estimate **est)
296 for (i = 0; i < m->size1; i++)
298 for (j = 0; j < m->size2; j++)
300 tmp = gsl_matrix_get (m, i, j) - est[j]->mean;
301 gsl_matrix_set (m, i, j, tmp);
305 struct innovations_estimate **
306 pspp_innovations (const struct design_matrix *dm, size_t lag)
308 struct innovations_estimate **est;
311 est = xnmalloc (dm->m->size2, sizeof *est);
312 for (i = 0; i < dm->m->size2; i++)
314 est[i] = xmalloc (sizeof *est[i]);
315 /* est[i]->variable = vars[i]; */
316 innovations_struct_init (est[i], dm, lag);
319 get_mean (dm->m, est);
320 subtract_mean (dm->m, est);
321 get_covariance (dm->m, est, lag);
322 get_coef (dm->m, est, lag);
328 pspp_innovations_free_one (struct innovations_estimate *est)
332 assert (est != NULL);
333 for (i = 0; i < (size_t) est->max_lag; i++)
335 pspp_coeff_free (est->coeff[i]);
342 void pspp_innovations_free (struct innovations_estimate **est, size_t n)
346 assert (est != NULL);
347 for (i = 0; i < n; i++)
349 pspp_innovations_free_one (est[i]);