ATSW — A Time Series Workshop

The Box-Jenkins-Treadway suite: exact maximum-likelihood estimation of univariate models, transfer functions and VARMA, with three MCP assistants so the whole methodology can be run by conversation.

pip install atsw

The ladder

The suite is one process, and the order is not a convention — it is a dependency. Each rung produces the input the next one needs.

rung	tool	you get
1	art (35 tools)	identification: λ, d, D, the seasonality frequency by frequency, the orders, the interventions → an .inp
2	fue	exact-ML estimation and the diagnosis → an .out and a .pre
3	mtram (21 tools)	transfer functions and networks: a directional model with a gain
3′	sima (15 tools)	simultaneous VARMA: both equations, impulse responses, variance decomposition

Rungs 3 and 3′ are alternatives, not a sequence. Which one fits depends on what you are prepared to assume — the sima example ends on exactly that question.

And the univariate model is the measuring stick. If a multivariate model cannot beat its forecasts, the multivariate model is what needs rethinking.

The examples run

Both series the worked examples use ship inside the package, so every node can be reproduced:

import atsw
atsw.example_path("IPC_ES.csv")    # Spanish CPI, 216 obs, 2002-01…2019-12
atsw.example_path("WTI.csv")       # oil, same window

How to run them — the prompts to paste, and what to expect.

Where to start

• New to the suite — read the art worked example. It follows one real series through every decision node, twice: once for forecasting and once as the input to a transfer model. The nodes are the same and some of the decisions are not, which is the part hardest to learn from a manual.
• Building a transfer model — mtram.
• Building a VARMA — sima.
• Writing files by hand or by program — File formats. Read the first section before anything else: there are two .inp dialects in this suite and they are not interchangeable.

Per-package documentation

package	what it documents
art-tseries	quickstart, the 35 art tools, architecture
drtran	the school's practice, the decision nodes, the ladder as an optimisation, the 21 mtram tools
drvarma	user and developer guides, the multivariate .inp, the 15 sima tools

Every tool reference is generated from the docstrings. In an MCP server the docstring is what the model reads, so the page you see and the instruction the assistant receives are the same text by construction.

Methodology

The analysis is not the canonical Box-Jenkins one but the version extended by Arthur B. Treadway, a disciple of Gwilym Jenkins, shaped by his work on the Forecasting and Monitoring Services of the Spanish economy. Box-Jenkins is a case of false simplicity: the ARMA models are simple, but the iterative building process worked wonderfully mainly if you were Box, Jenkins or one of their disciples. The hard part is the judgement, and that is what the assistants are for.