Linear Regression Test
Summary of Linear
Regression Certification Results for Xnumbers.xla, Matrix.xla, BigMatrix.xla
add-ins
These add-ins are not specific statistical
packages, but they contain a few useful
functions for linear regression and univariate summary statistic showing
interesting performance. Here, we report the NIST StRD test for Linear
Regression Coefficients applied to these add-ins and compared with MS Excel
|
Test |
Packages |
|
Excel 97/2000/2002 built-in function |
|
|
Excel 2003 built-in function |
|
|
Xnumbers.xla v 5.0
(multiprecision) |
|
|
Matrix.xla v 2.0 (standard
precision) |
|
|
BigMatrix.xla v 1.2 (fixed 35
digits precision) |
See the technical note NIST StRD
Benchmarks by Michael Kozluk for detailed description of these
test
|
NIST StRD Dataset Properties for Linear Regression |
|||||
|
Name |
Level of difficulty |
Model of class |
Paramet. |
Number of variables |
Points |
|
Norris |
low |
Linear |
2 |
1 |
36 |
|
Pontius |
low |
Quadratic |
3 |
1 |
40 |
|
NoInt1 |
medium |
Linear |
1 |
1 |
11 |
|
NoInt2 |
medium |
Linear |
1 |
1 |
3 |
|
Filip |
high |
Polynomial |
11 |
1 |
82 |
|
Longley |
high |
Multilinear |
7 |
6 |
16 |
|
Wampler1 |
high |
Polynomial |
6 |
1 |
21 |
|
Wampler2 |
high |
Polynomial |
6 |
1 |
21 |
|
Wampler3 |
high |
Polynomial |
6 |
1 |
21 |
|
Wampler4 |
high |
Polynomial |
6 |
1 |
21 |
|
Wampler5 |
high |
Polynomial |
6 |
1 |
21 |
In report comparisons, it is common use to
report the Log Relative Error (LRE) that puts in evidence the decimal digits of
accuracy. It cannot exceed 15 and higher values are obviously better.