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
|
Here's a checker program that runs libgretl against the NIST reference
datasets for linear regression, for details on which see
http://itl.nist.gov/div898/strd/general/main.html
The checker validates libgretl and (if available) the gretl multiple
precision plugin.
And there's a sub-dir named nist-nls with a rig for checking gretl's
nonlinear regression code against the NIST reference datasets.
NIST Results with libgretl
==========================
If all is well there should be no failures on the 11 NIST tests.
The checker program reads the "certified values" and data out of each
NIST file, runs a regression using libgretl, and compares the libgretl
estimates with the certified values, starting at a precision of 9
significant figures. If there is any disagreement, the comparison is
repeated using 8 significant figures, and so on. If the figures do
not agree at a precision of at least 6 figures, the test is deemed to
have been failed and an error message is printed.
You can get more details on gretl's performance on the tests by
running the 'nistcheck' program in verbose mode (-v flag) or very
verbose mode (-vv).
NIST Results with multiple precision plugin
===========================================
If the Gnu Multiple Precision library (GMP) was detected when gretl
was configured, and gretl's GMP plugin was built, then each NIST test
will be repeated using the plugin. All tests should give results
that agree with the NIST certified values to at least 12 significant
figures.
Note that the test data sets Wampler1 and Wampler2 give, by
construction, an exact polynomial fit. This is difficult to
reproduce. If you run the checker program in verbose mode you will
see that instead of standard errors of precisely zero, values around
10^{-39} are reported. You can, however, drive these to zero by
raising the precision used in the calculations. The default for
gretl's multiple precision plugin is to assign 256 bits for the
representation of each floating point number, which is enough for most
purposes. You can increase this value by setting the environment
variable GRETL_MP_BITS, for example:
GRETL_MP_BITS=4096 ./nistcheck -v
Allin Cottrell
last updated April 2011
|
