Skip to content

Migrating from the FUE in C

For people who already use Treadway's FUE. What is the same, what is called differently, and what genuinely differs in the output.

The short version: your .inp files work unchanged, the .out is the same layout, and the command line takes the same arguments. The differences are at the edges, and they are listed here rather than left to be discovered.

1. The command line

fue      model  [eml|aml] [chk|nochk] [-f [N]]      # the C program
fue-py   model  [eml|aml] [chk|nochk] [-f [N]]      # this package

Same arguments, same defaults (eml, chk, horizon 24), same convention of omitting the .inp extension. Likewise fuf → fuf-py.

⚠ Why the -py suffix, and why it will not be dropped. fue and fuf are the C programs and they live in /usr/local/bin; pip install --user puts its scripts in ~/.local/bin, which comes first in the PATH. Installing this package as fue would silently leave you running the port when you believe you are running the original — on the same file, in the same directory, with the same output name. The two coexist on purpose, and being able to compare them is the point of the whole exercise (docs/PORT.md).

C port note
eml / aml same exact / approximate ML
chk / nochk same check MA invertibility
-f [N] same write the forecast .inp for fuf
geom not implemented the geometric transform of fue.c has no counterpart yet

2. Files in, files out

In. The .inp is the same format, read by the same grammar — the parser reproduces the reading order of fue.c §3.0-3.7 (docs/FILE_CONTRACT.md). Since August 2026 the DRVUS-era files load too, without editing: the missing bands/refactor section is tolerated (bugs/BUG-0011) and a Latin-1 file is read as Latin-1 (bugs/BUG-0010). The nine Box-Jenkins specifications that ship with DRVUS 1.2.01 are in the test suite precisely so this keeps being true.

Out.

file C port
.out yes yes — same layout, diffable line by line
.pre disabled in fue-1.13.1 (preputf is commented out) yes
forecast .inp with -f yes yes
.tex, _dist.tex, _res.tex, .dvi yes not implemented — the port warns and skips
plots gnuplot matplotlib, from the Python API

The .pre is worth a paragraph. In the C it is commented out, and there is a reason recorded as bugs/BUG-0007: the C's .pre writer omitted easter, trend and non-standard variables, so fue C could not re-read its own .pre and segfaulted on it. The port writes a complete one, and the invariant that makes it useful is testable: run fue on a .pre and the numbers must not move (docs/FILE_CONTRACT.md §4).

3. What differs in the numbers

Nothing, to the precision that matters — and that claim is measured rather than asserted:

  • on the 28 preserved .out files of real work, the port reproduces the estimates, the termination code, the iteration count and the gradient norm;
  • on the nine Box-Jenkins specifications, log-likelihoods agree with the archived DRVUS runs to 1e-10 or better in eight of them.

The ninth is a1, and it is the one thing worth knowing before comparing an old .out with a new run: archived output from the 2000s was produced by 32-bit binaries carrying 80-bit intermediates. On a well-conditioned likelihood this is invisible; on a1 it decides the outcome, and Mauricio's own C rebuilt today gives what the port gives. docs/PROVENANCE.md §2.2, and tools/reproduce_drvus_reference.sh if you want to rebuild the old arithmetic and see it for yourself.

4. What the port adds

  • The optimiser's verdict reaches you. converged now means the gradient criterion was satisfied, not merely that nothing crashed; a fit that stopped because the iterates froze says so (docs/CONVERGENCE.md).
  • A library, not only a program. fue.Model(...).fit() — the .inp is one way in, not the only one.
  • It does not die on you. The C aborts the process on a degenerate series and on a failed eigensolve (bugs/BUG-0008, bugs/BUG-0009); the port raises.
  • A pure-Python engine when no compiled extension is available, running the same algorithm and the same optimiser.

5. What the port does not have

Said plainly, because a migration guide that omits this wastes your afternoon:

  • the LaTeX reports (.tex, _dist.tex, _res.tex) and the .dvi;
  • the geom transform;
  • the gnuplot graphics as such — there are matplotlib equivalents through the Python API, but they are not the same files.

6. A five-minute check before you switch

Treadway's advice for learning FUE was to run it on the shipped examples and compare with the shipped output; the same protocol works for the migration:

fue     model            # the original
fue-py  model            # this package
diff model.out model.out.new

If the two .out differ in more than the last digits, that is a finding and it should be reported with both files attached — which is exactly the protocol the manual asks for, and the one that produced most of bugs/.