Heval.jl is an AI-powered forecasting agent that combines LLMs with production-grade statistical models to automate time series analysis using natural language.
You provide time series data and a plain-English question. Heval automatically:
- Analyzes features (trend, seasonality, stationarity)
- Selects the best model via cross-validation (MASE, must beat SNaive)
- Generates forecasts with 80%/95% intervals
- Detects anomalies
- Explains results in natural language
It evaluates 16 models (ARIMA, ETS, BATS, TBATS, Theta, Croston, ARAR, Diffusion, and more) powered by Durbyn.jl.
Supports panel data, intermittent demand, flexible Tables.jl inputs, and local LLMs via Ollama (no API key required).
Note: The CI pipeline currently fails because Durbyn.jl is not yet in the Julia registry. Registration is expected in the next few weeks.
GitHub: GitHub - taf-society/Heval.jl
Early release (v0.1.0) – feedback and contributions welcome.