FreqAI: generalizing adaptive modeling for chaotic time-series market forecasts

A peer-reviewed software paper (submitted October 2022, published that December, reviewed by @ady00 and @shagunsodhani) for FreqAI, which “aims to provide a generalized and extensible open-sourced framework geared toward live deployments of adaptive modeling for market forecasting”, built on Freqtrade’s existing data collection, storage, and live-exchange interaction. The paper names the libraries it wraps — scikit-learn, CatBoost, LightGBM, XGBoost, stable_baselines3, OpenAI gym, TensorFlow, PyTorch, SciPy, NumPy, pandas — and describes a run as configured by two files, a --config and a --strategy. Its claim to contain methods “not available anywhere else in the open-source (or scientific) world” is the authors’ own.

The 2022 feature list has since been overtaken by the software. Current documentation describes self-adaptive retraining during live deployment, feature engineering over 10k+ features, backtesting with automated retraining, outlier removal, PCA dimensionality reduction, and model persistence to disk. Reinforcement learning is supported through stable_baselines3 and gym, with a templated ReinforcementLearner and three environments spanning hold/long/short, four-action, and five-action spaces.

The code is not a research drop: it lives under freqtrade/freqai/ in the maintained Freqtrade repository, which is GPL-3.0 — copyleft, so anything built on it inherits those terms. The paper text itself is CC BY 4.0, and the Zenodo archive record’s cc-by-4.0 metadata describes that record, not the software. The paper reports no returns, PnL, or profitability figures of any kind.