foreBlocks Docs

v0.1.15 — Modular Forecasting

Forecasting, preprocessing, search, and tooling in one repo

foreblocks is the forecasting library. foretools is the companion toolbox. Start from the stable training path, branch into preprocessing or transformers, and only then move into heavier workflows like DARTS search, conformal intervals, or the visual Web UI.

Minimal installCore stack with optional extras layered on top.

Stable pathForecastingModel + Trainer + ModelEvaluator.

NASDARTS with zero-cost screening, bilevel search, and retraining.

Uncertainty10 conformal methods — static, adaptive, and attention-based.

Start here

If you are new to the project, this is the safest reading order:

1. Overview
2. Getting Started
3. Public API
4. The subsystem guide that matches your workflow

Train a baseline first

Use the smallest reliable path through ForecastingModel, Trainer, and NumPy-backed dataloaders.

Start from raw series

Use TimeSeriesHandler when you need scaling, filtering, imputation, and window generation from a [T, D] array.

Transformer and MoE

Transformer, attention, patching, and MoE guides cover the more configurable internals.

Hybrid Mamba (SSM)

Pure SSM and hybrid SSM+attention blocks with Triton/CUDA kernel backends.

Search architectures

DARTS staged NAS pipeline: zero-cost ranking, bilevel search, and result analysis.

Prediction intervals

Post-hoc conformal prediction — 10 methods from split conformal to online ACI variants.

Documentation map

TutorialsGuidesForetoolsArchitecture

Step-by-step runnable examples:

• Getting Started
• Train A Direct Model
• Run A DARTS Search
• Generate Synthetic Series
• Optimize With BOHB

Subsystem deep dives:

• Preprocessor
• Custom Blocks
• Transformer
• Mixture of Experts
• Hybrid Mamba
• DARTS
• Evaluation & Metrics
• Uncertainty Quantification
• Web UI
• Troubleshooting

Companion utilities:

• Foretools Overview
• Time Series Generator
• BOHB Search
• VMD Decomposition
• AutoDA Augmentation

Internals and design notes:

• System Overview
• Forecasting Pipeline
• DARTS Search Pipeline

Stable public entry points

Start from the top-level exports before reaching for deep imports:

CoreArchitecture searchUncertaintyHybrid Mamba

from foreblocks import (
    ForecastingModel,
    Trainer,
    ModelEvaluator,
    TimeSeriesHandler,
    TimeSeriesDataset,
    create_dataloaders,
    ModelConfig,
    TrainingConfig,
)

from foreblocks.darts import (
    DARTSTrainer,
    DARTSConfig,
    DARTSTrainConfig,
    FinalTrainConfig,
    MultiFildelitySearchConfig,
)

from foreblocks.core import ConformalPredictionEngine

from foreblocks.hybrid_mamba import (
    HybridMambaBlock,
    HybridMamba2Block,
    TinyHybridMamba2LM,
)

The docs mirror the packaging split: the default install stays lean, while preprocessing, DARTS, MLTracker, VMD, and other heavier workflows are documented with the exact extras that enable them.

Notes

• The canonical docs source lives in docs/.
• The published versioned docs live under /docs/.
• The static landing page at site root lives in web/.
• foretools is documented here too, even when utilities are not re-exported from top-level foreblocks.