drvec 0.1.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- drvec/__init__.py +22 -0
- drvec/_gsl.py +83 -0
- drvec/api.py +942 -0
- drvec/c_engine.py +391 -0
- drvec/cast.py +196 -0
- drvec/cli.py +592 -0
- drvec/data.py +295 -0
- drvec/diagnose.py +439 -0
- drvec/fdhess.py +130 -0
- drvec/fit.py +528 -0
- drvec/forecast.py +314 -0
- drvec/irfboot.py +216 -0
- drvec/levels.py +158 -0
- drvec/loglik.py +45 -0
- drvec/nla.py +266 -0
- drvec/parallel.py +86 -0
- drvec/params.py +311 -0
- drvec/rank.py +401 -0
- drvec/report.py +1712 -0
- drvec/results.py +167 -0
- drvec/rng.py +51 -0
- drvec/rungtable.py +82 -0
- drvec/start.py +317 -0
- drvec/stderr.py +33 -0
- drvec/xlink.py +226 -0
- drvec-0.1.0.dist-info/METADATA +205 -0
- drvec-0.1.0.dist-info/RECORD +31 -0
- drvec-0.1.0.dist-info/WHEEL +5 -0
- drvec-0.1.0.dist-info/entry_points.txt +2 -0
- drvec-0.1.0.dist-info/licenses/COPYING +339 -0
- drvec-0.1.0.dist-info/top_level.txt +1 -0
drvec/data.py
ADDED
|
@@ -0,0 +1,295 @@
|
|
|
1
|
+
"""The `.pre` route: one fue model per series, read and cast by FUE.
|
|
2
|
+
|
|
3
|
+
C: `read_pre_inputs` (drvec.c:6777-6965) and `build_y2_levels` (696-765).
|
|
4
|
+
|
|
5
|
+
The reader and the univariate cast are fue's, as in drtran-python: `fue.load()`
|
|
6
|
+
reads each file and `fue.cast_us.cast_us_py` builds its series, at the .pre's
|
|
7
|
+
own parameters. drvec differs from drtran only in how it EMBEDS the model:
|
|
8
|
+
drtran takes fue's stationary w (step [7] of the cast, the non-stationary
|
|
9
|
+
operator applied); drvec works in LEVELS (it differences the nabla-Y2 block
|
|
10
|
+
itself, rank by rank), so it takes fue's cast of the same model with the
|
|
11
|
+
non-stationary operator emptied (`levels_cast`): step [6]'s
|
|
12
|
+
|
|
13
|
+
z = refactor * BoxCox(y) - det, det = sum of nu(B) I_t
|
|
14
|
+
|
|
15
|
+
(fue's regressor: `_build_indicator` on the file's own calendar and
|
|
16
|
+
`calcnu_py`'s impulse response), and nothing of step [7].
|
|
17
|
+
|
|
18
|
+
What is taken from each file: the series, its calendar, lambda, the rescaling
|
|
19
|
+
factor, the mean's value and flag (-case pre) and z. What is NOT: the
|
|
20
|
+
univariate ARMA (on this route drvec re-estimates the diagonal rung itself,
|
|
21
|
+
docs/PORT.md §4).
|
|
22
|
+
|
|
23
|
+
The component follows fue's rule, the suite's: nu(B) truncated at 40 lags for
|
|
24
|
+
a rational term (delta != 0), the variables summed and subtracted once (atsw-gui
|
|
25
|
+
lib/fuedet; fue >= 0.1.18). The oracle, drvec C 0.13, applies the same rule, so
|
|
26
|
+
w is the C's bit for bit; BUG-40, the C's old exact filter, is closed by
|
|
27
|
+
agreeing with fue (docs/PORT.md §13).
|
|
28
|
+
|
|
29
|
+
The deterministic component is kept over the common sample AND the forecast
|
|
30
|
+
horizon for the back-transform of the levels (bt_det, BUG-36): `Data.det_path`,
|
|
31
|
+
fue's `_build_xi` (the regressor fue's own forecast adds back), as drtran's
|
|
32
|
+
`to_level` does.
|
|
33
|
+
|
|
34
|
+
The files are lined up by DATE, on the intersection of their samples (the C's
|
|
35
|
+
rule, 6810-6840), which is not drvarma's ladder rule (same last date).
|
|
36
|
+
|
|
37
|
+
Refused: what fue's reader refuses, with fue's message (a type it does not
|
|
38
|
+
know, a malformed file); a NON-STANDARD deterministic variable (fue reads it as
|
|
39
|
+
a data column, the C refuses it by design: it is a transfer input, drtran's
|
|
40
|
+
job), with the C's message; `easter` off monthly data (fue would build a null
|
|
41
|
+
regressor, the C refuses it); and seasonal differences D > 0 (the C reads them
|
|
42
|
+
and silently applies none).
|
|
43
|
+
"""
|
|
44
|
+
from __future__ import annotations
|
|
45
|
+
|
|
46
|
+
from dataclasses import dataclass, field
|
|
47
|
+
|
|
48
|
+
import numpy as np
|
|
49
|
+
|
|
50
|
+
|
|
51
|
+
def abs_period(year, sub, freq):
|
|
52
|
+
"""drvec.c:6768: one index on a common calendar."""
|
|
53
|
+
if freq <= 1:
|
|
54
|
+
return int(year)
|
|
55
|
+
return int(year) * freq + (int(sub) - 1)
|
|
56
|
+
|
|
57
|
+
|
|
58
|
+
@dataclass
|
|
59
|
+
class RankData:
|
|
60
|
+
"""datamat and Y2_levels for one rank (build_y2_levels, levels layout).
|
|
61
|
+
|
|
62
|
+
`dm[t][i]` (0-based t = 0..n-1) is nabla Y2_i for i < s and Y1 in levels for
|
|
63
|
+
i >= s; `y2[t][i]` is the Y2 level. Both as lists (for the C-order loops of
|
|
64
|
+
the seeds) and as arrays (for the column-wise W of the cast)."""
|
|
65
|
+
r: int
|
|
66
|
+
n: int
|
|
67
|
+
dm: list
|
|
68
|
+
y2: list
|
|
69
|
+
dm_np: np.ndarray
|
|
70
|
+
y2_np: np.ndarray
|
|
71
|
+
|
|
72
|
+
|
|
73
|
+
@dataclass
|
|
74
|
+
class Data:
|
|
75
|
+
"""The system's LEVELS in w units, (nobs_raw, M), in the .inp's column
|
|
76
|
+
order: the first s series are the nabla-Y2 block, the last r the Y1 block."""
|
|
77
|
+
raw: np.ndarray
|
|
78
|
+
names: list
|
|
79
|
+
freq: int = 1
|
|
80
|
+
start: tuple = (1, 1)
|
|
81
|
+
files: list = field(default_factory=list)
|
|
82
|
+
refactors: list = field(default_factory=list)
|
|
83
|
+
pre_mean_free: list = field(default_factory=list) # each .pre frees its mean (Imu)
|
|
84
|
+
pre_mu: list = field(default_factory=list) # and its value
|
|
85
|
+
xl: object = None # -xpre/-xlink: xlink.Inputs (the inputs, with presample)
|
|
86
|
+
lambdas: list = field(default_factory=list) # each series' Box-Cox lambda
|
|
87
|
+
estwin: int = None # -estwin E: the fit sees rows 1..E of nabla Y (drvec.c:740)
|
|
88
|
+
det_src: list = field(default_factory=list) # per series: (fue Model, off) or None
|
|
89
|
+
_ranks: dict = field(default_factory=dict, repr=False)
|
|
90
|
+
ladder_cache: dict = field(default_factory=dict, repr=False)
|
|
91
|
+
|
|
92
|
+
@property
|
|
93
|
+
def M(self):
|
|
94
|
+
return self.raw.shape[1]
|
|
95
|
+
|
|
96
|
+
@property
|
|
97
|
+
def nobs_raw(self):
|
|
98
|
+
return self.raw.shape[0]
|
|
99
|
+
|
|
100
|
+
def with_raw(self, raw):
|
|
101
|
+
"""The same system on other levels (a bootstrap replication, a rescale):
|
|
102
|
+
a fresh object, so its rank data and its ladder cache start empty."""
|
|
103
|
+
return Data(np.asarray(raw, float), list(self.names), self.freq,
|
|
104
|
+
self.start, list(self.files), list(self.refactors),
|
|
105
|
+
list(self.pre_mean_free), list(self.pre_mu),
|
|
106
|
+
lambdas=list(self.lambdas), det_src=list(self.det_src))
|
|
107
|
+
|
|
108
|
+
def det_path(self, i, T):
|
|
109
|
+
"""bt_det[i][1..T] (drvec.c:6876-6882): series i's deterministic
|
|
110
|
+
component on the common sample and beyond, in w units, 0-based (row t
|
|
111
|
+
is raw observation t + 1). Zeros when the file declares none."""
|
|
112
|
+
src = self.det_src[i] if i < len(self.det_src) else None
|
|
113
|
+
if src is None:
|
|
114
|
+
return [0.0] * T
|
|
115
|
+
model, off = src
|
|
116
|
+
return det_path(model, off, T)
|
|
117
|
+
|
|
118
|
+
@property
|
|
119
|
+
def nobs(self):
|
|
120
|
+
"""The C's nobs: the rows of nabla Y the fit sees (-estwin trims it)."""
|
|
121
|
+
return self.estwin if self.estwin else self.nobs_raw - 1
|
|
122
|
+
|
|
123
|
+
def at(self, r, full=False):
|
|
124
|
+
"""build_y2_levels for rank r (cached); full=True ignores -estwin (the
|
|
125
|
+
rolling evaluation filters the whole sample)."""
|
|
126
|
+
key = r if (full is False or not self.estwin) else (r, "full")
|
|
127
|
+
if key in self._ranks:
|
|
128
|
+
return self._ranks[key]
|
|
129
|
+
M = self.M
|
|
130
|
+
s = M - r
|
|
131
|
+
n = self.nobs_raw - 1 if (full or not self.estwin) else self.estwin
|
|
132
|
+
if n < 3:
|
|
133
|
+
raise ValueError("the levels layout needs at least 4 observations")
|
|
134
|
+
raw = self.raw
|
|
135
|
+
dm = [[0.0] * M for _ in range(n)]
|
|
136
|
+
y2 = [[0.0] * s for _ in range(n)]
|
|
137
|
+
for t in range(n):
|
|
138
|
+
for i in range(s):
|
|
139
|
+
dm[t][i] = float(raw[t + 1, i]) - float(raw[t, i])
|
|
140
|
+
y2[t][i] = float(raw[t + 1, i])
|
|
141
|
+
for j in range(r):
|
|
142
|
+
dm[t][s + j] = float(raw[t + 1, s + j])
|
|
143
|
+
rd = RankData(r, n, dm, y2, np.array(dm, float),
|
|
144
|
+
np.array(y2, float) if s else np.zeros((n, 0)))
|
|
145
|
+
rd.xl = self.xl
|
|
146
|
+
rd.nobs_raw = self.nobs_raw
|
|
147
|
+
self._ranks[key] = rd
|
|
148
|
+
return rd
|
|
149
|
+
|
|
150
|
+
def first_estimated(self):
|
|
151
|
+
"""(year, sub) of the first observation of the estimation sample
|
|
152
|
+
(BUG-37: one after the raw start, the first is used by the difference)."""
|
|
153
|
+
y, sub = self.start
|
|
154
|
+
# nobs_raw - nobs + 1 (drvec.c:10970): with -estwin the C counts the
|
|
155
|
+
# window's trim at the START, so its date is late by nobs_full - E
|
|
156
|
+
k = abs_period(y, sub, self.freq) + (self.nobs_raw - self.nobs)
|
|
157
|
+
if self.freq <= 1:
|
|
158
|
+
return k, 1
|
|
159
|
+
return k // self.freq, k % self.freq + 1
|
|
160
|
+
|
|
161
|
+
|
|
162
|
+
|
|
163
|
+
|
|
164
|
+
NONSTANDARD_MSG = (
|
|
165
|
+
" Las variables NO ESTÁNDAR no se admiten por diseño: son\n"
|
|
166
|
+
" una versión rudimentaria de un modelo de transferencia con\n"
|
|
167
|
+
" input X. Especifica esa relación como transferencia (ω/δ, b),\n"
|
|
168
|
+
" que es precisamente lo que drtran estima.")
|
|
169
|
+
|
|
170
|
+
|
|
171
|
+
def load_model(path, what="ERROR: cannot read"):
|
|
172
|
+
"""fue.load(path) -> (ts, model), its refusals passed on with the file named.
|
|
173
|
+
A non-standard deterministic variable (fue's `custom`) is refused as the C
|
|
174
|
+
refuses it."""
|
|
175
|
+
import fue
|
|
176
|
+
try:
|
|
177
|
+
ts, model = fue.load(str(path))
|
|
178
|
+
except (ValueError, IndexError, KeyError, OSError) as e:
|
|
179
|
+
raise ValueError(f"{e}\n{what} {path}") from None
|
|
180
|
+
freq = int(ts.freq) if ts.freq else 1
|
|
181
|
+
for k, itv in enumerate(model.interventions):
|
|
182
|
+
if itv.type == "custom":
|
|
183
|
+
raise ValueError(f"{path}: variable determinista {k + 1} no reconocida.\n"
|
|
184
|
+
+ NONSTANDARD_MSG)
|
|
185
|
+
if itv.type == "easter" and freq != 12:
|
|
186
|
+
# fue reads it and builds an identically null regressor (its
|
|
187
|
+
# _build_indicator generates easter on monthly data only); the C
|
|
188
|
+
# refuses it. Refused rather than carried as a term that is zero.
|
|
189
|
+
raise ValueError(f"{path}: variable determinista {k + 1} (easter) on data of "
|
|
190
|
+
f"frequency {freq}: easter is monthly only")
|
|
191
|
+
return ts, model
|
|
192
|
+
|
|
193
|
+
|
|
194
|
+
def levels_cast(model):
|
|
195
|
+
"""fue's cast of `model` at its own parameters, with the non-stationary
|
|
196
|
+
operator emptied: -> (phi, theta, mu, z), z[0..nobs-1] = refactor *
|
|
197
|
+
BoxCox(y) - det (cast_us_py's step [6]; step [7] is the identity).
|
|
198
|
+
|
|
199
|
+
The model with d = D = 0 and no ifadf factors is built as drtran builds the
|
|
200
|
+
input re-differenced by its output's operator (_alt_est_specs): a copy of
|
|
201
|
+
fue's model with only the operator changed, `build_est_spec`, and the cast
|
|
202
|
+
at `_build_initial_x` (the .pre's values in fue's own parameter order).
|
|
203
|
+
phi and theta are fue's factors multiplied out (`_unscramble`, the
|
|
204
|
+
fixed-frequency AR(2)/MA(2) included), mu the mean's value whatever its
|
|
205
|
+
flag."""
|
|
206
|
+
import copy
|
|
207
|
+
from fue.cast_us import _build_initial_x, build_est_spec, cast_us_py
|
|
208
|
+
alt = copy.deepcopy(model)
|
|
209
|
+
alt.d, alt.D = 0, 0
|
|
210
|
+
alt.ifadf = [0] * len(model.ifadf) if model.ifadf else model.ifadf
|
|
211
|
+
spec = build_est_spec(alt)
|
|
212
|
+
_p, _q, phi, theta, mu, z, ifault = cast_us_py(_build_initial_x(alt), spec)
|
|
213
|
+
if ifault:
|
|
214
|
+
raise ValueError("a fixed-frequency AR(2)/MA(2) factor of the .pre has c2 >= 0: "
|
|
215
|
+
"fue's cast refuses it")
|
|
216
|
+
return [float(v) for v in phi], [float(v) for v in theta], float(mu), z
|
|
217
|
+
|
|
218
|
+
|
|
219
|
+
def det_path(model, off, T):
|
|
220
|
+
"""The deterministic component of fue's `model` on the file's rows off+1 ..
|
|
221
|
+
off+T (its own calendar, t = 1 its first observation), beyond the sample if
|
|
222
|
+
needed: fue's `_build_xi` at the .pre's omegas and deltas (a step or a
|
|
223
|
+
trend goes on). -> list of T floats."""
|
|
224
|
+
from fue.forecast import _build_xi
|
|
225
|
+
nobs = int(model.series.nobs)
|
|
226
|
+
freq = int(model.series.freq) if model.series.freq else 1
|
|
227
|
+
horizon = max(0, off + T - nobs)
|
|
228
|
+
xi = _build_xi(model, nobs, freq, horizon,
|
|
229
|
+
[list(i.omega) for i in model.interventions],
|
|
230
|
+
[list(i.delta) for i in model.interventions])
|
|
231
|
+
return [float(v) for v in xi[1 + off:1 + off + T]]
|
|
232
|
+
|
|
233
|
+
|
|
234
|
+
def _check_positive(f, ts, lam):
|
|
235
|
+
if abs(lam - 1.0) > 1.0e-12:
|
|
236
|
+
for t, z in enumerate(ts.data):
|
|
237
|
+
if float(z) <= 0.0:
|
|
238
|
+
raise ValueError(f"{f} asks for a Box-Cox with lambda = {lam:g} and "
|
|
239
|
+
f"observation {t + 1} is {float(z):g}")
|
|
240
|
+
|
|
241
|
+
|
|
242
|
+
def read_pre(files):
|
|
243
|
+
"""Read M fue files and line them up by date (read_pre_inputs)."""
|
|
244
|
+
if len(files) < 2:
|
|
245
|
+
raise ValueError(f"a VEC model needs at least two series, got {len(files)}")
|
|
246
|
+
series = []
|
|
247
|
+
freq0 = None
|
|
248
|
+
first = last = None
|
|
249
|
+
for k, f in enumerate(files):
|
|
250
|
+
ts, model = load_model(f)
|
|
251
|
+
if ts.nobs < 4:
|
|
252
|
+
raise ValueError(f"{f} carries {ts.nobs} observations")
|
|
253
|
+
freq = int(ts.freq) if ts.freq else 1
|
|
254
|
+
if freq0 is None:
|
|
255
|
+
freq0 = freq
|
|
256
|
+
elif freq != freq0:
|
|
257
|
+
raise ValueError(f"{f} has frequency {freq} and {files[0]} has {freq0}. "
|
|
258
|
+
"A VEC model is one\n system on one calendar; the "
|
|
259
|
+
"series cannot be mixed.")
|
|
260
|
+
if int(model.D or 0) > 0:
|
|
261
|
+
raise NotImplementedError(
|
|
262
|
+
f"{f} declares D = {model.D}: drvec differences only the nabla-Y2 "
|
|
263
|
+
"block, regularly; the C ignores D in silence, the port refuses it")
|
|
264
|
+
y, sub = ts.start
|
|
265
|
+
st = abs_period(y, sub if freq > 1 else 1, freq)
|
|
266
|
+
en = st + int(ts.nobs) - 1
|
|
267
|
+
first = st if first is None else max(first, st)
|
|
268
|
+
last = en if last is None else min(last, en)
|
|
269
|
+
series.append((f, ts, model, st))
|
|
270
|
+
if last - first + 1 < 4:
|
|
271
|
+
raise ValueError(f"the .pre files overlap in {last - first + 1} observation(s); "
|
|
272
|
+
"at least 4 are\n needed. They are lined up by DATE, not "
|
|
273
|
+
"by position.")
|
|
274
|
+
n = last - first + 1
|
|
275
|
+
M = len(series)
|
|
276
|
+
raw = np.zeros((n, M))
|
|
277
|
+
names, refs, mfree, mus, lams, dets = [], [], [], [], [], []
|
|
278
|
+
for i, (f, ts, model, st) in enumerate(series):
|
|
279
|
+
off = first - st
|
|
280
|
+
lam = float(model.boxlam)
|
|
281
|
+
_check_positive(f, ts, lam)
|
|
282
|
+
names.append(ts.name or str(f))
|
|
283
|
+
refs.append(float(model.refactor))
|
|
284
|
+
lams.append(lam)
|
|
285
|
+
mfree.append(bool(model.estimate_mu))
|
|
286
|
+
mus.append(float(model.mu0))
|
|
287
|
+
_phi, _theta, _mu, z = levels_cast(model)
|
|
288
|
+
dets.append((model, off) if model.interventions else None)
|
|
289
|
+
raw[:, i] = z[off:off + n]
|
|
290
|
+
if freq0 <= 1:
|
|
291
|
+
start = (first, 1)
|
|
292
|
+
else:
|
|
293
|
+
start = (first // freq0, first % freq0 + 1)
|
|
294
|
+
return Data(raw, names, freq0, start, [str(f) for f in files], refs, mfree, mus,
|
|
295
|
+
lambdas=lams, det_src=dets)
|