---

name: forecasting-experiments description: "Construct and adapt TimeMixer forecasting experiment commands." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Forecasting Experiments

Use this sub-skill when the user wants to construct, adapt, audit, or explain TimeMixer commands for forecasting experiments. It covers long-term forecasting and short-term forecasting presets for ETT, electricity/ECL, Traffic, Weather, Solar, PEMS, and M4, plus custom forecasting command construction.

Route here for

• Building a safe run.py command for a known forecast benchmark preset.
• Adapting a benchmark command to a custom CSV or changed forecast horizon.
• Explaining training, validation, testing, checkpoints, printed metrics, test_results/, and M4 forecast CSV outputs.
• Reducing epochs, batch size, GPU usage, or worker count before a user-approved expensive run.
• Debugging forecast-specific CLI, data-dimension, GPU, checkpoint, and M4 evaluation failures.

Route elsewhere

• Raw dataset file layout, CSV columns, benchmark download placement, and deterministic data validation: use data-preparation.
• TimeMixer model internals, tensor shapes inside PDM/FMM, decomposition choices, and forward smoke checks: use model-architecture.
• Imputation, anomaly detection, or classification commands: use universal-tasks.

Bundled operating assets

• references/benchmark-recipes.md lists source-script-derived forecasting presets and marks full benchmark execution as external-data/expensive.
• references/workflows.md explains command-building, training/test-only behavior, checkpoints, result locations, and custom adaptation.
• references/troubleshooting.md covers common forecast failures, including the run.py --help percent-format bug and CPU fallback.
• scripts/build_timemixer_command.py prints a run.py command or JSON plan; it never launches training.

Safe default procedure

1. Identify the forecast family: long-term tabular/array data, PEMS traffic, Solar, or M4 seasonal short-term forecasting.
2. Use the bundled command builder to print, inspect, and adjust the command before any execution.
3. Confirm external data availability and user approval before running a full benchmark-scale training command.
4. For custom data, ensure enc_in, dec_in, and c_out match the selected feature mode and validated data dimensions.
5. Treat original benchmark shell scripts as distilled evidence only; do not invoke them from this skill.