Journal of Open Source Software JOSS 2475-9066 Open Journals 0 N/A FreqAI: generalizing adaptive modeling for chaotic time-series market forecasts 0000-0001-5618-8629 Ph.D Robert A. Caulk 0000-0003-3289-8604 Ph.D Elin Törnquist Voppichler Matthias Lawless Andrew R. McMullan Ryan Santos Wagner Costa Pogue Timothy C. van der Vlugt Johan Gehring Stefan P. Schmidt Pascal Emergent Methods LLC, Arvada Colorado, 80005, USA Freqtrade open source project ¿VOL? ¿ISSUE? ¿PAGE? Authors of papers retain copyright and release the work under a Creative Commons Attribution 4.0 International License (CC BY 4.0) 2022 The article authors Authors of papers retain copyright and release the work under a Creative Commons Attribution 4.0 International License (CC BY 4.0) Python Machine Learning adaptive modeling chaotic systems time-series forecasting Statement of need

Forecasting chaotic time-series based systems, such as equity/cryptocurrency markets, requires a broad set of tools geared toward testing a wide range of hypotheses. Fortunately, a recent maturation of robust machine learning libraries (e.g. scikit-learn), has opened up a wide range of research possibilities. Scientists from a diverse range of fields can now easily prototype their studies on an abundance of established machine learning algorithms. Similarly, these user-friendly libraries enable “citzen scientists” to use their basic Python skills for data-exploration. However, leveraging these machine learning libraries on historical and live chaotic data sources can be logistically difficult and expensive. Additionally, robust data-collection, storage, and handling presents a disparate challenge. FreqAI aims to provide a generalized and extensible open-sourced framework geared toward live deployments of adaptive modeling for market forecasting. The FreqAI framework is effectively a sandbox for the rich world of open-source machine learning libraries. Inside the FreqAI sandbox, users find they can combine a wide variety of third-party libraries to test creative hypotheses on a free live 24/7 chaotic data source - cryptocurrency exchange data.

Summary

FreqAI evolved from a desire to test and compare a range of adaptive time-series forecasting methods on chaotic data. Cryptocurrency markets provide a unique data source since they are operational 24/7 and the data is freely available. Luckily, an existing open-source software, Freqtrade, had already matured under a range of talented developers to support robust data collection/storage, as well as robust live environmental interactions for standard algorithmic trading. Freqtrade also provides a set of data analysis/visualization tools for the evaluation of historical performance as well as live environmental feedback. FreqAI builds on top of Freqtrade to include a user-friendly well tested interface for integrating external machine learning libraries for adaptive time-series forecasting. Beyond enabling the integration of existing libraries, FreqAI hosts a range of custom algorithms and methodologies aimed at improving computational and predictive performances. Thus, FreqAI contains a range of unique features which can be easily tested in combination with all the existing Python-accessible machine learning libraries to generate novel research on live and historical data.

The high-level overview of the software is depicted in Figure 1.

freqai-algo Abstracted overview of FreqAI algorithm

Connecting machine learning libraries

Although the FreqAI framework is designed to accommodate any Python library in the “Model training” and “Feature set engineering” portions of the software (Figure 1), it already boasts a wide range of well documented examples based on various combinations of:

scikit-learn (Pedregosa et al., 2011), Catboost (Prokhorenkova et al., 2018), LightGBM (Ke et al., 2017), XGBoost (Chen & Guestrin, 2016), stable_baselines3 (Raffin et al., 2021), openai gym (Brockman et al., 2016), tensorflow (Abadi et al., 2015), pytorch (Paszke et al., 2019), Scipy (Virtanen et al., 2020), Numpy (Harris et al., 2020), and pandas (McKinney & others, 2010).

These mature projects contain a wide range of peer-reviewed and industry standard methods, including:

Regression, Classification, Neural Networks, Reinforcement Learning, Support Vector Machines, Principal Component Analysis, point clustering, and much more.

which are all leveraged in FreqAI for users to use as templates or extend with their own methods.

Furnishing novel methods and features

Beyond the industry standard methods available through external libraries - FreqAI includes novel methods which are not available anywhere else in the open-source (or scientific) world. For example, FreqAI provides :

a custom algorithm/methodology for adaptive modeling

rapid and self-monitored feature engineering tools

unique model features/indicators

optimized data collection algorithms

safely integrated outlier detection methods

websocket communicated forecasts

Of particular interest for researchers, FreqAI provides the option of large scale experimentation via an optimized websocket communications interface.

Optimizing the back-end

FreqAI aims to make it simple for users to combine all the above tools to run studies based in two distinct modules:

backtesting studies

live-deployments

Both of these modules and their respective data management systems are built on top of Freqtrade, a mature and actively developed cryptocurrency trading software. This means that FreqAI benefits from a wide range of tangential/disparate feature developments such as:

FreqUI, a graphical interface for backtesting and live monitoring

telegram control

robust database handling

futures/leverage trading

dollar cost averaging

trading strategy handling

a variety of free data sources via CCXT (FTX, Binance, Kucoin etc.)

These features derive from a strong external developer community that shares in the benefit and stability of a communal CI (Continuous Integration) system. Beyond the developer community, FreqAI benefits strongly from the userbase of Freqtrade, where most FreqAI beta-testers/developers originated. This symbiotic relationship between Freqtrade and FreqAI ignited a thoroughly tested beta, which demanded a four month beta and comprehensive documentation containing:

numerous example scripts

a full parameter table

methodological descriptions

high-resolution diagrams/figures

detailed parameter setting recommendations

Providing a reproducible foundation for researchers

FreqAI provides an extensible, robust, framework for researchers and citizen data scientists. The FreqAI sandbox enables rapid conception and testing of exotic hypotheses. From a research perspective, FreqAI handles the multitude of logistics associated with live deployments, historical backtesting, and feature engineering. With FreqAI, researchers can focus on their primary interests of feature engineering and hypothesis testing rather than figuring out how to collect and handle data. Further - the well maintained and easily installed open-source framework of FreqAI enables reproducible scientific studies. This reproducibility component is essential to general scientific advancement in time-series forecasting for chaotic systems.

Technical details

Typical users configure FreqAI via two files:

A configuration file (--config) which provides access to the full parameter list available here:

control high-level feature engineering

customize adaptive modeling techniques

set any model training parameters available in third-party libraries

manage adaptive modeling parameters (retrain frequency, training window size, continual learning, etc.)

A strategy file (--strategy) where users:

list of the base training features

set standard technical-analysis strategies

control trade entry/exit criteria

With these two files, most users can exploit a wide range of pre-existing integrations in Catboost and 7 other libraries with a simple command:

freqtrade trade --config config_freqai.example.json --strategy FreqaiExampleStrategy --freqaimodel CatboostRegressor

Advanced users will edit one of the existing --freqaimodel files, which are simply an children of the IFreqaiModel (details below). Within these files, advanced users can customize training procedures, prediction procedures, outlier detection methods, data preparation, data saving methods, etc. This is all configured in a way where they can customize as little or as much as they want. This flexible customization is owed to the foundational architecture in FreqAI, which is comprised of three distinct Python objects:

IFreqaiModel

A singular long-lived object containing all the necessary logic to collect data, store data, process data, engineer features, run training, and inference models.

FreqaiDataKitchen

A short-lived object which is uniquely created for each asset/model. Beyond metadata, it also contains a variety of data processing tools.

FreqaiDataDrawer

Singular long-lived object containing all the historical predictions, models, and save/load methods.

These objects interact with one another with one goal in mind - to provide a clean data set to machine learning experts/enthusiasts at the user endpoint. These power-users interact with an inherited IFreqaiModel that allows them to dig as deep or as shallow as they wish into the inheritence tree. Typical power-users focus their efforts on customizing training procedures and testing exotic functionalities available in third-party libraries. Thus, power-users are freed from the algorithmic weight associated with data management, and can instead focus their energy on testing creative hypotheses. Meanwhile, some users choose to override deeper functionalities within IFreqaiModel to help them craft unique data structures and training procedures.

The class structure and algorithmic details are depicted in the following diagram:

image Class diagram summarizing object interactions in FreqAI

Online documentation

The documentation for FreqAI is available online at https://www.freqtrade.io/en/latest/freqai/ and covers a wide range of materials:

Quick-start with a single command and example files - (beginners)

Introduction to the feature engineering interface and basic configurations - (intermediate users)

Parameter table with indepth descriptions and default parameter setting recommendations - (intermediate users)

Data analysis and post-processing - (advanced users)

Methodological considerations complemented by high resolution figures - (advanced users)

Instructions for integrating third party machine learning libraries into custom prediction models - (advanced users)

Software architectural description with class diagram - (developers)

File structure descriptions - (developers)

The docs direct users to a variety of pre-made examples which integrate Catboost, LightGBM, XGBoost, Sklearn, stable_baselines3, torch, tensorflow. Meanwhile, developers will also find thorough docstrings and type hinting throughout the source code to aid in code readability and customization.

FreqAI also benefits from a strong support network of users and developers on the Freqtrade discord as well as on the FreqAI discord. Within the FreqAI discord, users will find a deep and easily searched knowledge base containing common errors. But more importantly, users in the FreqAI discord share anectdotal and quantitative observations which compare performance between various third-party libraries and methods.

State of the field

There are two other open-source tools which are geared toward helping users build models for time-series forecasts on market based data. However, each of these tools suffer from a non-generalized frameworks that do not permit comparison of methods and libraries. Additionally, they do not permit easy live-deployments or adaptive-modeling methods. For example, two open-sourced projects called tensortrade (Tensortrade, 2022) and FinRL (AI4Finance-Foundation, 2022) limit users to the exploration of reinforcement learning on historical data. These softwares also do not provide robust live deployments, they do not furnish novel feature engineering algorithms, and they do not provide custom data analysis tools. FreqAI fills the gap.

On-going research

Emergent Methods, based in Arvada CO, is actively using FreqAI to perform large scale experiments aimed at comparing machine learning libraries in live and historical environments. Past projects include backtesting parametric sweeps, while active projects include a 3 week live deployment comparison between CatboosRegressor, LightGBMRegressor, and XGBoostRegressor. Results from these studies are on track for publication in scientific journals as well as more general data science blogs (e.g. Medium).

Installing and running <monospace>FreqAI</monospace>

FreqAI is automatically installed with Freqtrade using the following commands on linux systems:

git clone git@github.com:freqtrade/freqtrade.git cd freqtrade ./setup.sh -i

However, FreqAI also benefits from Freqtrade docker distributions, and can be run with docker by pulling the stable or develop images from Freqtrade distributions.

Funding sources

FreqAI has had no official sponsors, and is entirely grass roots. All donations into the project (e.g. the GitHub sponsor system) are kept inside the project to help support development of open-sourced and communally beneficial features.

Acknowledgements

We would like to acknowledge various beta testers of FreqAI:

Richárd Józsa

Juha Nykänen

Salah Lamkadem

As well as various Freqtrade developers maintaining tangential, yet essential, modules.

PedregosaF. VaroquauxG. GramfortA. MichelV. ThirionB. GriselO. BlondelM. PrettenhoferP. WeissR. DubourgV. VanderplasJ. PassosA. CournapeauD. BrucherM. PerrotM. DuchesnayE. Scikit-learn: Machine learning in Python Journal of Machine Learning Research 2011 12 2825 2830 ProkhorenkovaLiudmila GusevGleb VorobevAleksandr DorogushAnna Veronika GulinAndrey CatBoost: Unbiased boosting with categorical features Proceedings of the 32nd international conference on neural information processing systems Curran Associates Inc. Red Hook, NY, USA 2018 6639 6649 KeGuolin MengQi FinleyThomas WangTaifeng ChenWei MaWeidong YeQiwei LiuTie-Yan Lightgbm: A highly efficient gradient boosting decision tree Advances in neural information processing systems 2017 30 3146 3154 ChenTianqi GuestrinCarlos XGBoost: A scalable tree boosting system Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining ACM New York, NY, USA 2016 978-1-4503-4232-2 http://doi.acm.org/10.1145/2939672.2939785 10.1145/2939672.2939785 785 794 RaffinAntonin HillAshley GleaveAdam KanervistoAnssi ErnestusMaximilian DormannNoah Stable-Baselines3: Reliable reinforcement learning implementations Journal of Machine Learning Research 2021 22 268 http://jmlr.org/papers/v22/20-1364.html 1 8 BrockmanGreg CheungVicki PetterssonLudwig SchneiderJonas SchulmanJohn TangJie ZarembaWojciech OpenAI gym 2016 https://arxiv.org/abs/1606.01540 AbadiMartín AgarwalAshish BarhamPaul BrevdoEugene ChenZhifeng CitroCraig CorradoGreg S. DavisAndy DeanJeffrey DevinMatthieu GhemawatSanjay GoodfellowIan HarpAndrew IrvingGeoffrey IsardMichael JiaYangqing JozefowiczRafal KaiserLukasz KudlurManjunath LevenbergJosh ManéDandelion MongaRajat MooreSherry MurrayDerek OlahChris SchusterMike ShlensJonathon SteinerBenoit SutskeverIlya TalwarKunal TuckerPaul VanhouckeVincent VasudevanVijay ViégasFernanda VinyalsOriol WardenPete WattenbergMartin WickeMartin YuYuan ZhengXiaoqiang TensorFlow: Large-scale machine learning on heterogeneous systems 2015 https://www.tensorflow.org/ PaszkeAdam GrossSam MassaFrancisco LererAdam BradburyJames ChananGregory KilleenTrevor LinZeming GimelsheinNatalia AntigaLuca DesmaisonAlban KopfAndreas YangEdward DeVitoZachary RaisonMartin TejaniAlykhan ChilamkurthySasank SteinerBenoit FangLu BaiJunjie ChintalaSoumith PyTorch: An imperative style, high-performance deep learning library Advances in neural information processing systems 32 WallachH. LarochelleH. BeygelzimerA. dAlché-BucF. FoxE. GarnettR. Curran Associates, Inc. 2019 http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf 8024 8035 VirtanenPauli GommersRalf OliphantTravis E. HaberlandMatt ReddyTyler CournapeauDavid BurovskiEvgeni PetersonPearu WeckesserWarren BrightJonathan van der WaltStéfan J. BrettMatthew WilsonJoshua MillmanK. Jarrod MayorovNikolay NelsonAndrew R. J. JonesEric KernRobert LarsonEric CareyC J Polatİlhan FengYu MooreEric W. VanderPlasJake LaxaldeDenis PerktoldJosef CimrmanRobert HenriksenIan QuinteroE. A. HarrisCharles R. ArchibaldAnne M. RibeiroAntônio H. PedregosaFabian van MulbregtPaul SciPy 1.0 Contributors SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python Nature Methods 2020 17 10.1038/s41592-019-0686-2 261 272 HarrisCharles R. MillmanK. Jarrod WaltStéfan J. van der GommersRalf VirtanenPauli CournapeauDavid WieserEric TaylorJulian BergSebastian SmithNathaniel J. KernRobert PicusMatti HoyerStephan KerkwijkMarten H. van BrettMatthew HaldaneAllan RíoJaime Fernández del WiebeMark PetersonPearu Gérard-MarchantPierre SheppardKevin ReddyTyler WeckesserWarren AbbasiHameer GohlkeChristoph OliphantTravis E. Array programming with NumPy Nature Springer Science; Business Media LLC 202009 585 7825 https://doi.org/10.1038/s41586-020-2649-2 10.1038/s41586-020-2649-2 357 362 McKinneyWes others Data structures for statistical computing in python Proceedings of the 9th python in science conference Austin, TX 2010 445 51 56 AI4Finance-foundation 2022 20220930 https://github.com/AI4Finance-Foundation/FinRL Tensortrade 2022 20220930 https://tensortradex.readthedocs.io/en/latest/L