Running the examples yourself¶

The three worked examples are not illustrations — they are analyses of two real series, and both series ship inside the package so you can rerun every node and get the same numbers.

pip install atsw

>>> import atsw
>>> atsw.example_path("IPC_ES.csv")   # Spanish CPI,  216 obs, 2002-01…2019-12
>>> atsw.example_path("WTI.csv")      # WTI oil spot, same window
>>> atsw.example_path()               # the directory, with PROVENANCE.md

Both are public statistics (INE and EIA/FRED). They stop in 2019-12 on purpose: 2020 onwards contains the COVID collapse and the 2022 energy shock, and those two level breaks turn a lesson about identification into a lesson about interventions.

The normal way: ask the assistant¶

This suite is driven by conversation. You do not run a script; you tell Claude what you want and it drives the tools, showing you the evidence at each node and waiting for your decision. Register the servers once:

claude mcp add art   -- art-mcp
claude mcp add mtram -- mtram
claude mcp add sima  -- sima

Then paste one of these.

The `art` example — identification, node by node¶

Load atsw's example series IPC_ES.csv (you can find it with python -c "import atsw; print(atsw.example_path('IPC_ES.csv'))"). Take me through the GUIDED identification, stopping at every node: the Box-Cox, the order of integration, the seasonality frequency by frequency, the ARMA orders and the interventions. At each one show me the evidence, tell me what you would decide and why, and give me the argument against your own choice before I confirm.

That last sentence is the one that matters. The tools produce the evidence; the value of the assistant is the argument, and asking for the counter-argument is how you stop it anchoring you.

Compare what you get with WORKED_EXAMPLE_ART.md, Part I. Your decisions may differ from the recorded ones — at the f=3 frontier they legitimately can — and that is the point of a guided pipeline rather than a button.

The same series, for a different purpose¶

Now do it again, but the model is going to be the OUTPUT of a transfer model whose input is the oil price. Which of the decisions change, and why?

Two of them do. Part II of the same document says which and gives the reason.

The `mtram` example — the pass-through¶

Using atsw's example series, build a transfer model of IPC_ES on WTI: first the univariate models with art and fue, then load both .pre files in mtram, check the operators, identify the link from the prewhitened CCF and estimate. Show me the diagonal gate before anything else, and tell me what the gain means in economic terms.

Expect a gain near 0.027: a permanent 1 % move in oil passes through to about 0.027 % of the Spanish CPI level.

The `sima` example — the same pair, simultaneously¶

Take the same two series and fit a bivariate VAR with sima — logs, one regular difference, and deseasonalise. Give me the impulse responses and the variance decomposition, and tell me what the Cholesky ordering assumes and what changes if I reverse it.

The reversal is the point. With a contemporaneous residual correlation near 0.5, how the shared variance is attributed is decided by the ordering, not by the data.

And do not skip the deseasonalising. Without it that correlation comes out at 0.23 instead of 0.51 and the decomposition changes accordingly — measured in Compared with statsmodels, which also shows that statsmodels.VARMAX has no seasonal handling at all.

Without an assistant¶

Every server is also an ordinary Python library, and the CLIs work on their own:

fue MODEL              # estimate an .inp, write .out and .pre
drtran-py Y.pre X.pre -b 0 -r 0 -s 1     # a transfer model
drvarma DATA.inp -p 1                    # a VARMA

You will get the same numbers. What you will not get is the part the examples are actually about — which question each node is answering, and why one criterion rather than another. That is what the assistant adds, and it is also what a colleague would add.

If your numbers differ¶

Three usual causes, in order of frequency:

A different window. The examples run 2002-01…2019-12. Including 2020+ changes every diagnostic.
A different seasonal specification. Deterministic harmonics and ∇₁₂ are not two spellings of one model; on this very series they disagree about the pass-through by a factor of five once the operators stop matching. See FILE_FORMATS.md on ifadf, and WORKED_EXAMPLE_MTRAM.md on the operators.
A decision you took differently. Which is allowed, and is why the documents record the argument and not only the answer.