Context parroting

Context parroting and time series foundation models for zero-shot forecasting of dynamical systems.

Context parroting is an informative baseline for time series foundation models. It generates naive forecasts by copying contiguous sequences of context points.

Repository layout

Superseded material lives in benchmark/legacy/ and analysis/legacy/, each with a README explaining what it is and what replaced it.

Datasets

The dysts zero-shot forecasting dataset was studied in earlier work. To avoid data fragmentation, we point to the original source files associated with that paper.

data/ includes:

Benchmarks

The unified driver runs any model in the registry against the full dysts dataset:

# Run all models
python benchmark/run_dysts_benchmarks.py

# Run specific models at a single context length
python benchmark/run_dysts_benchmarks.py chronos timemoe --context-lengths 512

# Split one slow model across parallel processes
python benchmark/run_dysts_benchmarks.py chronos --equations Lorenz Rossler --output-suffix _chunk0

# Available models: arima, chronos, chronos_bolt, dynamix, moirai2, panda_patchtst,
#                   parrot, simplex, timemoe, timesfm (2.0, paper version), timesfm_2p5

Results are written to benchmark/benchmark_results/<model>_dysts/ as {average,median}_{vpt,smape,vpt_2,smape_2,cdim,kl,mse,mae}.npy, each a dict keyed by (equation_name, context_length). benchmark/forecast_models.py is the model registry — add a new model by subclassing ForecastModel and registering it. DynaMix additionally requires cloning DynaMix-python into benchmark/DynaMix/.

The benchmarks can also be run with uv:

uv run python benchmark/run_dysts_benchmarks.py

Conventions (important for comparing numbers)

• sMAPE is on the 0–200 scale (predicting the mean of white noise gives ~200), defined locally in benchmark/metrics.py. It is deliberately not imported from dysts, whose default switched to a 0–100 scale in April 2025 — numbers produced against the wrong convention are silently half-sized (see analysis/legacy/README.md for a cautionary example).
• Initial conditions: for each (equation, context_length) pair the driver samples num_ic (default 20) starts uniformly across the full trajectory with a deterministic per-pair seed, so runs are reproducible and unbiased.

Contents of benchmark/

Reference

Information about the experiments can be found in the accompanying paper.

@inproceedings{zhang2026context,
  title={Context parroting: A simple but tough-to-beat baseline for foundation models in scientific machine learning},
  author={Yuanzhao Zhang and William Gilpin},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026}
}