AutoFit-TS

AutoFit-TS is a research-grade framework for learning and auditing temporal foundation models on irregular, long-horizon longitudinal data. It is designed for high-integrity scientific evaluation: reproducible benchmarks, strict no-leakage validation, and concept-level explanations that support auditability in downstream analyses.

Key features

• Automatic composition of temporal architectures to match dataset characteristics (irregular sampling, missingness, non-stationarity, multiscale rhythms, and exogenous influence) under a fixed compute/budget constraint.
• Multi-stream temporal backbone with modular exogenous conditioning (cross- attention, FiLM, bridge-token fusion) and interchangeable foundation wrappers.
• Concept Bottleneck layer that produces interpretable narrative indices that enable additive concept models and structured natural-language explanation reports.
• Parquet-native data ingestion and diagnostics, end-to-end benchmark pipelines, and automated generation of paper-ready evaluation tables (main, ablation, faithfulness, efficiency).

Repository layout

• src/narrative/ — core pipeline, model components, diagnostics, and explainability
• scripts/ — utilities for data preparation, benchmarking, auditing, and report generation
• configs/ — canonical experiment/dataset configurations
• docs/ — methodological notes, API references, and evaluation descriptors
• runs/ — runtime artifacts and generated reports (not tracked in the canonical release)

Getting started (developer quickstart)

1. Create a Python environment (recommended: Python 3.9+).
2. Install runtime dependencies from pyproject.toml.
3. Run unit tests:

PYTHONPATH=src pytest -q

Example: small-scale smoke benchmark

# build a small offers parquet (implementation-specific)
# run a minimal benchmark with limit rows and a short horizon
PYTHONPATH=src python scripts/run_full_benchmark.py --offers_path data/raw/offers --limit_rows 5000 --label_horizon 3 --models dlinear patchtst

Citation If you use this code in research, please cite the AutoFit-TS project and include a reference to the accompanying paper (when available).

License and contribution This repository is intended as a public research artifact. See LICENSE for terms. Contributions are welcome via issues and pull requests; please follow the repository's contribution guidelines.