redrvfl 0.1.0

RVFL, Ensemble Deep RVFL, and Recurrent Ensemble Deep RVFL models for regression and forecasting.

RedRVFL

RedRVFL is a lightweight Python package for financial time-series forecasting with Random Vector Functional Link models:

• RVFLRegressor: single hidden-layer RVFL with ridge readout.
• EDRVFLRegressor: ensemble deep RVFL with independent layer readouts.
• REDRVFLRegressor: recurrent ensemble deep RVFL for ordered time-series frames.
• Hyperopt/TPE tuning for RVFL.
• Layerwise Hyperopt/TPE tuning for EDRVFL and REDRVFL.

Installation

Core install:

git clone https://github.com/statsdl/RedRVFL.git
cd RedRVFL
pip install .

Finance experiment dependencies:

pip install ".[finance]"

For development:

pip install -e ".[dev]"
pytest

DJI Paper-Style Experiment

from redrvfl.finance import download_dji, run_dji_paper_experiment

download_dji("datasets/DJI.csv")
results = run_dji_paper_experiment(
    dataset_path="datasets/DJI.csv",
    seeds=(0,),
    horizon=20,
    look_ahead=1,
    n_layers=10,
    max_evals=100,
)

The split follows the paper: 70% training, 10% validation, and 20% test in chronological order. Hyperparameters are selected on validation data, then the model is fitted on train+validation and evaluated on the final test segment.

Command-line usage:

python examples/run_dji_paper.py --download --seeds 0 --horizon 20 --look-ahead 1
python examples/run_dji_paper.py --seeds 0,1,2

Hyperopt Tuning

from hyperopt import hp
from redrvfl.tuning import layerwise_tune_redrvfl

result = layerwise_tune_redrvfl(
    X,
    y,
    n_layers=10,
    layer_space={
        "n_hidden": hp.quniform("n_hidden", 20, 200, 1),
        "regularization": hp.uniform("regularization", 0, 1),
        "input_scale": hp.uniform("input_scale", 0, 1),
    },
    fixed_params={
        "recurrent_scale": 0.1,
        "random_state": 0,
    },
    validation_fraction=0.1 / 0.8,
    max_evals=100,
)

API

Supported package code lives in src/redrvfl.

All estimators expose:

• fit(X, y): train readout weights.
• predict(X): return predictions.
• predict(X, return_layers=True): for EDRVFL/REDRVFL, return each layer's prediction.

make_forecasting_frame(series, order, horizon) converts a time series into supervised lagged samples.

Finance and tuning utilities are intentionally imported from submodules:

• redrvfl.tuning
• redrvfl.finance

This keeps import redrvfl lightweight and avoids importing optional Hyperopt and Yahoo Finance dependencies unless they are needed.

Repository Notes

The legacy/, utils/, RecRVFL_/, and ForecastLib.py files are retained for traceability to earlier research scripts. They are not included in the published wheel and are not the supported package API.

PyPI Release

The publish workflow uses PyPI Trusted Publishing. Configure the PyPI trusted publisher with:

• owner: statsdl
• repository: RedRVFL
• workflow: publish.yml
• environment: pypi

License

MIT