Documentation website

Название исследуемой задачи:	Epiplexity and transfer in synthetic time-series pretraining
Тип научной работы:	CoIS, BMM
Авторы:	Danila Chernousov, Mark Ikonnikov, Daniil Kazachkov, Alexey Kravatskiy
Научный руководитель:	уточняется
Научный консультант(при наличии):	уточняется

Abstract

The practical goal is to train a time-series foundation model on high-quality synthetic data and select productive generators for that data. A productive generator supplies pretraining data that improve forecasting on real datasets under a fixed training-data budget. We study epiplexity-based measurements as possible signals for this selection. The implemented episynth package generates forecasting tasks, trains a probabilistic Moirai transformer with pre-update loss logging and EMA weights, estimates signed excess prequential loss, and evaluates transfer on GIFT-Eval. Saved experiments cover eleven priors at approximately 2 million and 40 million training windows, plus one additional Waveform initialization. KernelSynth transfers best at both budgets; the full-set association between excess area and aggregate CRPS is weak. The technical report documents the evidence, endpoint sensitivity, forecast failures, and limits of interpreting the area as epiplexity.

The earlier research roadmap in publications/plan/ is outdated. Current implementation details are in src/episynth/README.md and publications/description.md; the report documents the saved experiments.

Technical report and measurement methods

Read the technical report PDF and its reproduction instructions. The proposed four-method library scope is described in the measurement contracts:

1. Prequential coding (implemented): measure the predictive cost of the training sequence; return total, residual-data and signed model-proxy costs in bits.
2. Requential coding (registered stub): describe a student's learned predictions through teacher-selected synthetic training samples and their coding costs.
3. Predictive V-information (proposed): fit predictors with and without context; estimate the held-out log-loss reduction in bits per window.
4. Bayesian predictive log-loss (OU reference implemented): marginalize the known AR(1) prior to obtain joint and pointwise forecast-loss references. A common public wrapper remains proposed.

Seeded task generation, AdamW/EMA training, and GIFT-Eval scoring provide shared infrastructure. The saved transfer experiments use the prequential method.

Research publications

Presentations at conferences on the topic of research

Software modules developed as part of the study

1. The python package episynth with all implementation here.
2. Runnable demo notebook here.

Repository layout

src/episynth/      the package: generators, task windows, Moirai port, training with per-window
                   loss logging, analysis, evaluation on GIFT-Eval, pluggable estimators of
                   epiplexity (overview in src/episynth/README.md)
tests/             pytest
experiments/       experiment code and demo (demos/) plus saved outputs (episynth/): two training
                   series (configs, training curves, held-out losses, GIFT-Eval summaries,
                   analysis tables, figures, scripts)
doc/               Sphinx documentation (autodoc); GitHub Pages publication is manual
publications/      non-code material for the article and project: paper/ (arXiv manuscript),
                   blogpost/, notes/ (literature review), research_notes/, slides/, plan/,
                   figures/, technical_report/ (article and reproducible analysis), journal/ (research
                   journal of the time-series track: landscape and experiment, journal.pdf), description.md
pyproject.toml     package metadata (name = episynth)
README.rst         this file

Development

Requires Python 3.10+ and uv. On a network that needs system TLS certificates, add --system-certs to the uv commands.

uv sync --locked --extra dev
uv run --no-sync pytest tests/
uv run --no-sync ruff check src/ tests/ experiments/demos/ doc/source/conf.py

Documentation

The Sphinx documentation (with autodoc) is built on pushes and pull requests to main. Automatic publication is disabled as required by the course. To update GitHub Pages, open Actions → docs → Run workflow, select main, and start the workflow manually. To build it locally:

uv run --no-sync sphinx-build -b html -W --keep-going doc/source doc/_build
# open doc/_build/index.html

The generated API example is doc/_build/api/estimators.html. See doc/source/installation.rst for standalone hosting and GitHub Pages setup. The Pages source must be configured as GitHub Actions. The expected URL is https://intsystems.github.io/Epiplexity/; a successful deployment is required to update the online site. Pull requests build documentation without publishing.