Crate anofox_forecast 

anofox-forecast

Time series forecasting library for Rust.

Provides 35+ forecasting models including ARIMA, ETS, Theta, and baseline methods, along with seasonality decomposition (STL/MSTL), changepoint detection, and outlier detection.

For comprehensive periodicity detection, see the fdars crate.

Architecture Decisions

Cross-Validation Split (ts_cv_split)

Time series cross-validation with data leakage prevention is implemented in the forecast-extension DuckDB extension rather than in this crate. This section documents the rationale.

Why CV Split is Not Part of TimeSeries

The TimeSeries struct represents a single time series with its values, timestamps, and metadata. Cross-validation splitting was considered as a method on TimeSeries but was intentionally kept separate for these reasons:

1. Cross-series coordination: Fold generation is a global operation across multiple series, not a per-series operation. CV requires consistent fold boundaries across all series in a dataset.

2. External feature handling: Unknown future features like stockout flags or segment_id changes are external columns that don’t belong in the series data model. These require schema-aware handling at the data layer.

3. Data manipulation efficiency: DuckDB’s vectorized execution is more efficient for the bulk data operations (filtering, joining, filling) that CV split requires.

4. Schema flexibility: SQL macros can handle arbitrary column schemas without requiring Rust to know the schema at compile time.

Component Distribution

Component	Location	Rationale
Fold generation	DuckDB extension	Cross-series coordination, global operation
Train/test assignment	SQL/DuckDB	Simple comparison, vectorized execution
Unknown feature filling	Rust UDF via DuckDB	Per-series state tracking
Orchestration	SQL macro	Flexible, schema-agnostic

Using CV Functionality

For time series cross-validation with data leakage prevention, use the ts_cv_split function from the forecast-extension:

-- Example: Generate CV folds with unknown feature handling
SELECT * FROM ts_cv_split(
    my_data,
    n_splits := 3,
    horizon := 7,
    unknown_features := ['stockout', 'segment_id']
);

See forecast-extension#54 for implementation details.

Future Considerations

If per-series CV semantics become necessary in Rust (e.g., for standalone use without DuckDB), the fold generation logic could be extracted:

pub struct CvFoldGenerator {
    n_splits: usize,
    horizon: usize,
    gap: usize,
}

impl CvFoldGenerator {
    pub fn folds(&self, series_len: usize) -> Vec<usize> {
        // Returns training end indices for each fold
    }
}

This would allow fold generation to be shared while keeping data manipulation in the appropriate layer (SQL for multi-series datasets, Rust for single-series use).

Re-exports

pub use error::ForecastError;

pub use error::Result;

Modules

changepoint

Changepoint detection algorithms.

core

Core data structures for time series forecasting.

detection

Detection utilities for time series analysis.

error

Error types for the anofox-forecast library.

features

Time series feature extraction.

models

Forecasting models.

postprocess

Probabilistic forecasting via postprocessing.

prelude

seasonality

Seasonality detection and decomposition.

simd

SIMD-accelerated primitives via Trueno (f32 internal).

transform

Data transformations for time series.

utils

Utility functions for forecasting models.

validation

Statistical validation tests for time series models.