fix footer

.github/workflows/draft-pdf.yml

@@ -0,0 +1,23 @@
+on: [push]
+
+jobs:
+  paper:
+    runs-on: ubuntu-latest
+    name: Paper Draft
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v2
+      - name: Build draft PDF
+        uses: openjournals/openjournals-draft-action@master
+        with:
+          journal: joss
+          # This should be the path to the paper within your repo.
+          paper-path: docs/JOSS_paper/paper.md
+      - name: Upload
+        uses: actions/upload-artifact@v1
+        with:
+          name: paper
+          # This is the output path where Pandoc will write the compiled
+          # PDF. Note, this should be the same directory as the input
+          # paper.md
+          path: docs/JOSS_paper/paper.pdf

docs/JOSS_paper/note_to_editors.txt

@@ -0,0 +1,15 @@
+Dear Editors,
+We present a paper for ``FreqAI`` a machine learning sandbox for researchers and citizen scientists alike.
+There are a large number of authors, however all have contributed in a significant way to this paper.
+For clarity the contribution of each author is outlined:
+
+  - Robert Caulk            : Conception and software development
+  - Elin Tornquist          : Theoretical brainstorming, data analysis, tool dev
+  - Matthias Voppichler     : Software architecture and code review
+  - Andrew R. Lawless       : Extensive testing, feature brainstorming
+  - Ryan McMullan           : Extensive testing, feature brainstorming
+  - Wagner Costa Santos     : Major backtesting developments, extensive testing
+  - Pascal Schmidt          : Extensive testing, feature brainstorming
+  - Timothy C. Pogue        : Webhooks forecast sharing
+  - Stefan P. Gehring       : Extensive testing, feature brainstorming
+  - Johan van der Vlugt     : Extensive testing, feature brainstorming

docs/JOSS_paper/paper.bib

@@ -0,0 +1,207 @@
+@article{scikit-learn,
+ title={Scikit-learn: Machine Learning in {P}ython},
+ author={Pedregosa, F. and Varoquaux, G. and Gramfort, A. and Michel, V.
+         and Thirion, B. and Grisel, O. and Blondel, M. and Prettenhofer, P.
+         and Weiss, R. and Dubourg, V. and Vanderplas, J. and Passos, A. and
+         Cournapeau, D. and Brucher, M. and Perrot, M. and Duchesnay, E.},
+ journal={Journal of Machine Learning Research},
+ volume={12},
+ pages={2825--2830},
+ year={2011}
+}
+
+@inproceedings{catboost,
+author = {Prokhorenkova, Liudmila and Gusev, Gleb and Vorobev, Aleksandr and Dorogush, Anna Veronika and Gulin, Andrey},
+title = {CatBoost: Unbiased Boosting with Categorical Features},
+year = {2018},
+publisher = {Curran Associates Inc.},
+address = {Red Hook, NY, USA},
+abstract = {This paper presents the key algorithmic techniques behind CatBoost, a new gradient boosting toolkit. Their combination leads to CatBoost outperforming other publicly available boosting implementations in terms of quality on a variety of datasets. Two critical algorithmic advances introduced in CatBoost are the implementation of ordered boosting, a permutation-driven alternative to the classic algorithm, and an innovative algorithm for processing categorical features. Both techniques were created to fight a prediction shift caused by a special kind of target leakage present in all currently existing implementations of gradient boosting algorithms. In this paper, we provide a detailed analysis of this problem and demonstrate that proposed algorithms solve it effectively, leading to excellent empirical results.},
+booktitle = {Proceedings of the 32nd International Conference on Neural Information Processing Systems},
+pages = {6639–6649},
+numpages = {11},
+location = {Montr\'{e}al, Canada},
+series = {NIPS'18}
+}
+
+
+@article{lightgbm,
+  title={Lightgbm: A highly efficient gradient boosting decision tree},
+  author={Ke, Guolin and Meng, Qi and Finley, Thomas and Wang, Taifeng and Chen, Wei and Ma, Weidong and Ye, Qiwei and Liu, Tie-Yan},
+  journal={Advances in neural information processing systems},
+  volume={30},
+  pages={3146--3154},
+  year={2017}
+} 
+
+@inproceedings{xgboost,
+ author = {Chen, Tianqi and Guestrin, Carlos},
+ title = {{XGBoost}: A Scalable Tree Boosting System},
+ booktitle = {Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining},
+ series = {KDD '16},
+ year = {2016},
+ isbn = {978-1-4503-4232-2},
+ location = {San Francisco, California, USA},
+ pages = {785--794},
+ numpages = {10},
+ url = {http://doi.acm.org/10.1145/2939672.2939785},
+ doi = {10.1145/2939672.2939785},
+ acmid = {2939785},
+ publisher = {ACM},
+ address = {New York, NY, USA},
+ keywords = {large-scale machine learning},
+}
+
+@article{stable-baselines3,
+  author  = {Antonin Raffin and Ashley Hill and Adam Gleave and Anssi Kanervisto and Maximilian Ernestus and Noah Dormann},
+  title   = {Stable-Baselines3: Reliable Reinforcement Learning Implementations},
+  journal = {Journal of Machine Learning Research},
+  year    = {2021},
+  volume  = {22},
+  number  = {268},
+  pages   = {1-8},
+  url     = {http://jmlr.org/papers/v22/20-1364.html}
+}
+
+@misc{openai,
+      title={OpenAI Gym}, 
+      author={Greg Brockman and Vicki Cheung and Ludwig Pettersson and Jonas Schneider and John Schulman and Jie Tang and Wojciech Zaremba},
+      year={2016},
+      eprint={1606.01540},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG}
+}
+
+@misc{tensorflow,
+title={ {TensorFlow}: Large-Scale Machine Learning on Heterogeneous Systems},
+url={https://www.tensorflow.org/},
+note={Software available from tensorflow.org},
+author={
+    Mart\'{i}n~Abadi and
+    Ashish~Agarwal and
+    Paul~Barham and
+    Eugene~Brevdo and
+    Zhifeng~Chen and
+    Craig~Citro and
+    Greg~S.~Corrado and
+    Andy~Davis and
+    Jeffrey~Dean and
+    Matthieu~Devin and
+    Sanjay~Ghemawat and
+    Ian~Goodfellow and
+    Andrew~Harp and
+    Geoffrey~Irving and
+    Michael~Isard and
+    Yangqing Jia and
+    Rafal~Jozefowicz and
+    Lukasz~Kaiser and
+    Manjunath~Kudlur and
+    Josh~Levenberg and
+    Dandelion~Man\'{e} and
+    Rajat~Monga and
+    Sherry~Moore and
+    Derek~Murray and
+    Chris~Olah and
+    Mike~Schuster and
+    Jonathon~Shlens and
+    Benoit~Steiner and
+    Ilya~Sutskever and
+    Kunal~Talwar and
+    Paul~Tucker and
+    Vincent~Vanhoucke and
+    Vijay~Vasudevan and
+    Fernanda~Vi\'{e}gas and
+    Oriol~Vinyals and
+    Pete~Warden and
+    Martin~Wattenberg and
+    Martin~Wicke and
+    Yuan~Yu and
+    Xiaoqiang~Zheng},
+  year={2015},
+}
+
+@incollection{pytorch,
+title = {PyTorch: An Imperative Style, High-Performance Deep Learning Library},
+author = {Paszke, Adam and Gross, Sam and Massa, Francisco and Lerer, Adam and Bradbury, James and Chanan, Gregory and Killeen, Trevor and Lin, Zeming and Gimelshein, Natalia and Antiga, Luca and Desmaison, Alban and Kopf, Andreas and Yang, Edward and DeVito, Zachary and Raison, Martin and Tejani, Alykhan and Chilamkurthy, Sasank and Steiner, Benoit and Fang, Lu and Bai, Junjie and Chintala, Soumith},
+booktitle = {Advances in Neural Information Processing Systems 32},
+editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
+pages = {8024--8035},
+year = {2019},
+publisher = {Curran Associates, Inc.},
+url = {http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf}
+}
+
+@ARTICLE{scipy,
+  author  = {Virtanen, Pauli and Gommers, Ralf and Oliphant, Travis E. and
+            Haberland, Matt and Reddy, Tyler and Cournapeau, David and
+            Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and
+            Bright, Jonathan and {van der Walt}, St{\'e}fan J. and
+            Brett, Matthew and Wilson, Joshua and Millman, K. Jarrod and
+            Mayorov, Nikolay and Nelson, Andrew R. J. and Jones, Eric and
+            Kern, Robert and Larson, Eric and Carey, C J and
+            Polat, {\.I}lhan and Feng, Yu and Moore, Eric W. and
+            {VanderPlas}, Jake and Laxalde, Denis and Perktold, Josef and
+            Cimrman, Robert and Henriksen, Ian and Quintero, E. A. and
+            Harris, Charles R. and Archibald, Anne M. and
+            Ribeiro, Ant{\^o}nio H. and Pedregosa, Fabian and
+            {van Mulbregt}, Paul and {SciPy 1.0 Contributors}},
+  title   = {{{SciPy} 1.0: Fundamental Algorithms for Scientific
+            Computing in Python}},
+  journal = {Nature Methods},
+  year    = {2020},
+  volume  = {17},
+  pages   = {261--272},
+  adsurl  = {https://rdcu.be/b08Wh},
+  doi     = {10.1038/s41592-019-0686-2},
+}
+
+@Article{numpy,
+ title         = {Array programming with {NumPy}},
+ author        = {Charles R. Harris and K. Jarrod Millman and St{\'{e}}fan J.
+                 van der Walt and Ralf Gommers and Pauli Virtanen and David
+                 Cournapeau and Eric Wieser and Julian Taylor and Sebastian
+                 Berg and Nathaniel J. Smith and Robert Kern and Matti Picus
+                 and Stephan Hoyer and Marten H. van Kerkwijk and Matthew
+                 Brett and Allan Haldane and Jaime Fern{\'{a}}ndez del
+                 R{\'{i}}o and Mark Wiebe and Pearu Peterson and Pierre
+                 G{\'{e}}rard-Marchant and Kevin Sheppard and Tyler Reddy and
+                 Warren Weckesser and Hameer Abbasi and Christoph Gohlke and
+                 Travis E. Oliphant},
+ year          = {2020},
+ month         = sep,
+ journal       = {Nature},
+ volume        = {585},
+ number        = {7825},
+ pages         = {357--362},
+ doi           = {10.1038/s41586-020-2649-2},
+ publisher     = {Springer Science and Business Media {LLC}},
+ url           = {https://doi.org/10.1038/s41586-020-2649-2}
+}
+
+ @inproceedings{pandas,
+  title={Data structures for statistical computing in python},
+  author={McKinney, Wes and others},
+  booktitle={Proceedings of the 9th Python in Science Conference},
+  volume={445},
+  pages={51--56},
+  year={2010},
+  organization={Austin, TX},
+  doi={10.25080/Majora-92bf1922-00a}
+}
+
+
+
+@online{finrl,
+  title = {AI4Finance-Foundation},
+  year = 2022,
+  url = {https://github.com/AI4Finance-Foundation/FinRL},
+  urldate = {2022-09-30}
+}
+
+
+@online{tensortrade,
+  title = {tensortrade},
+  year = 2022,
+  url = {https://tensortradex.readthedocs.io/en/latest/L},
+  urldate = {2022-09-30}
+}

docs/JOSS_paper/paper.jats

@@ -0,0 +1,941 @@
+<?xml version="1.0" encoding="utf-8" ?>
+<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Publishing DTD v1.2 20190208//EN"
+                  "JATS-publishing1.dtd">
+<article xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink" dtd-version="1.2" article-type="other">
+<front>
+<journal-meta>
+<journal-id></journal-id>
+<journal-title-group>
+<journal-title>Journal of Open Source Software</journal-title>
+<abbrev-journal-title>JOSS</abbrev-journal-title>
+</journal-title-group>
+<issn publication-format="electronic">2475-9066</issn>
+<publisher>
+<publisher-name>Open Journals</publisher-name>
+</publisher>
+</journal-meta>
+<article-meta>
+<article-id pub-id-type="publisher-id">0</article-id>
+<article-id pub-id-type="doi">N/A</article-id>
+<title-group>
+<article-title><monospace>FreqAI</monospace>: generalizing adaptive
+modeling for chaotic time-series market forecasts</article-title>
+</title-group>
+<contrib-group>
+<contrib contrib-type="author">
+<contrib-id contrib-id-type="orcid">0000-0001-5618-8629</contrib-id>
+<name>
+<surname>Ph.D</surname>
+<given-names>Robert A. Caulk</given-names>
+</name>
+<xref ref-type="aff" rid="aff-1"/>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<contrib contrib-type="author">
+<contrib-id contrib-id-type="orcid">0000-0003-3289-8604</contrib-id>
+<name>
+<surname>Ph.D</surname>
+<given-names>Elin Törnquist</given-names>
+</name>
+<xref ref-type="aff" rid="aff-1"/>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<contrib contrib-type="author">
+<name>
+<surname>Voppichler</surname>
+<given-names>Matthias</given-names>
+</name>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<contrib contrib-type="author">
+<name>
+<surname>Lawless</surname>
+<given-names>Andrew R.</given-names>
+</name>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<contrib contrib-type="author">
+<name>
+<surname>McMullan</surname>
+<given-names>Ryan</given-names>
+</name>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<contrib contrib-type="author">
+<name>
+<surname>Santos</surname>
+<given-names>Wagner Costa</given-names>
+</name>
+<xref ref-type="aff" rid="aff-1"/>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<contrib contrib-type="author">
+<name>
+<surname>Pogue</surname>
+<given-names>Timothy C.</given-names>
+</name>
+<xref ref-type="aff" rid="aff-1"/>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<contrib contrib-type="author">
+<name>
+<surname>van der Vlugt</surname>
+<given-names>Johan</given-names>
+</name>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<contrib contrib-type="author">
+<name>
+<surname>Gehring</surname>
+<given-names>Stefan P.</given-names>
+</name>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<contrib contrib-type="author">
+<name>
+<surname>Schmidt</surname>
+<given-names>Pascal</given-names>
+</name>
+<xref ref-type="aff" rid="aff-2"/>
+</contrib>
+<aff id="aff-1">
+<institution-wrap>
+<institution>Emergent Methods LLC, Arvada Colorado, 80005,
+USA</institution>
+</institution-wrap>
+</aff>
+<aff id="aff-2">
+<institution-wrap>
+<institution>Freqtrade open source project</institution>
+</institution-wrap>
+</aff>
+</contrib-group>
+<volume>¿VOL?</volume>
+<issue>¿ISSUE?</issue>
+<fpage>¿PAGE?</fpage>
+<permissions>
+<copyright-statement>Authors of papers retain copyright and release the
+work under a Creative Commons Attribution 4.0 International License (CC
+BY 4.0)</copyright-statement>
+<copyright-year>2022</copyright-year>
+<copyright-holder>The article authors</copyright-holder>
+<license license-type="open-access" xlink:href="https://creativecommons.org/licenses/by/4.0/">
+<license-p>Authors of papers retain copyright and release the work under
+a Creative Commons Attribution 4.0 International License (CC BY
+4.0)</license-p>
+</license>
+</permissions>
+<kwd-group kwd-group-type="author">
+<kwd>Python</kwd>
+<kwd>Machine Learning</kwd>
+<kwd>adaptive modeling</kwd>
+<kwd>chaotic systems</kwd>
+<kwd>time-series forecasting</kwd>
+</kwd-group>
+</article-meta>
+</front>
+<body>
+<sec id="statement-of-need">
+  <title>Statement of need</title>
+  <p>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. <monospace>scikit-learn</monospace>), 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.
+  <ext-link ext-link-type="uri" xlink:href="https://www.freqtrade.io/en/latest/freqai/"><monospace>FreqAI</monospace></ext-link>
+  aims to provide a generalized and extensible open-sourced framework
+  geared toward live deployments of adaptive modeling for market
+  forecasting. The <monospace>FreqAI</monospace> framework is
+  effectively a sandbox for the rich world of open-source machine
+  learning libraries. Inside the <monospace>FreqAI</monospace> 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.</p>
+</sec>
+<sec id="summary">
+  <title>Summary</title>
+  <p><ext-link ext-link-type="uri" xlink:href="https://www.freqtrade.io/en/latest/freqai/"><monospace>FreqAI</monospace></ext-link>
+  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,
+  <ext-link ext-link-type="uri" xlink:href="https://www.freqtrade.io/en/stable/"><monospace>Freqtrade</monospace></ext-link>,
+  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.
+  <monospace>Freqtrade</monospace> also provides a set of data
+  analysis/visualization tools for the evaluation of historical
+  performance as well as live environmental feedback.
+  <monospace>FreqAI</monospace> builds on top of
+  <monospace>Freqtrade</monospace> 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, <monospace>FreqAI</monospace> hosts a range of
+  custom algorithms and methodologies aimed at improving computational
+  and predictive performances. Thus, <monospace>FreqAI</monospace>
+  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.</p>
+  <p>The high-level overview of the software is depicted in Figure
+  1.</p>
+  <p><named-content content-type="image">freqai-algo</named-content>
+  <italic>Abstracted overview of FreqAI algorithm</italic></p>
+  <sec id="connecting-machine-learning-libraries">
+    <title>Connecting machine learning libraries</title>
+    <p>Although the <monospace>FreqAI</monospace> 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:</p>
+    <list list-type="bullet">
+      <list-item>
+        <p>scikit-learn
+        (<xref alt="Pedregosa et al., 2011" rid="ref-scikit-learn" ref-type="bibr">Pedregosa
+        et al., 2011</xref>), Catboost
+        (<xref alt="Prokhorenkova et al., 2018" rid="ref-catboost" ref-type="bibr">Prokhorenkova
+        et al., 2018</xref>), LightGBM
+        (<xref alt="Ke et al., 2017" rid="ref-lightgbm" ref-type="bibr">Ke
+        et al., 2017</xref>), XGBoost
+        (<xref alt="Chen &amp; Guestrin, 2016" rid="ref-xgboost" ref-type="bibr">Chen
+        &amp; Guestrin, 2016</xref>), stable_baselines3
+        (<xref alt="Raffin et al., 2021" rid="ref-stable-baselines3" ref-type="bibr">Raffin
+        et al., 2021</xref>), openai gym
+        (<xref alt="Brockman et al., 2016" rid="ref-openai" ref-type="bibr">Brockman
+        et al., 2016</xref>), tensorflow
+        (<xref alt="Abadi et al., 2015" rid="ref-tensorflow" ref-type="bibr">Abadi
+        et al., 2015</xref>), pytorch
+        (<xref alt="Paszke et al., 2019" rid="ref-pytorch" ref-type="bibr">Paszke
+        et al., 2019</xref>), Scipy
+        (<xref alt="Virtanen et al., 2020" rid="ref-scipy" ref-type="bibr">Virtanen
+        et al., 2020</xref>), Numpy
+        (<xref alt="Harris et al., 2020" rid="ref-numpy" ref-type="bibr">Harris
+        et al., 2020</xref>), and pandas
+        (<xref alt="McKinney &amp; others, 2010" rid="ref-pandas" ref-type="bibr">McKinney
+        &amp; others, 2010</xref>).</p>
+      </list-item>
+    </list>
+    <p>These mature projects contain a wide range of peer-reviewed and
+    industry standard methods, including:</p>
+    <list list-type="bullet">
+      <list-item>
+        <p>Regression, Classification, Neural Networks, Reinforcement
+        Learning, Support Vector Machines, Principal Component Analysis,
+        point clustering, and much more.</p>
+      </list-item>
+    </list>
+    <p>which are all leveraged in <monospace>FreqAI</monospace> for
+    users to use as templates or extend with their own methods.</p>
+  </sec>
+  <sec id="furnishing-novel-methods-and-features">
+    <title>Furnishing novel methods and features</title>
+    <p>Beyond the industry standard methods available through external
+    libraries - <monospace>FreqAI</monospace> includes novel methods
+    which are not available anywhere else in the open-source (or
+    scientific) world. For example, <monospace>FreqAI</monospace>
+    provides :</p>
+    <list list-type="bullet">
+      <list-item>
+        <p>a custom algorithm/methodology for adaptive modeling</p>
+      </list-item>
+      <list-item>
+        <p>rapid and self-monitored feature engineering tools</p>
+      </list-item>
+      <list-item>
+        <p>unique model features/indicators</p>
+      </list-item>
+      <list-item>
+        <p>optimized data collection algorithms</p>
+      </list-item>
+      <list-item>
+        <p>safely integrated outlier detection methods</p>
+      </list-item>
+      <list-item>
+        <p>websocket communicated forecasts</p>
+      </list-item>
+    </list>
+    <p>Of particular interest for researchers,
+    <monospace>FreqAI</monospace> provides the option of large scale
+    experimentation via an optimized websocket communications
+    interface.</p>
+  </sec>
+  <sec id="optimizing-the-back-end">
+    <title>Optimizing the back-end</title>
+    <p><monospace>FreqAI</monospace> aims to make it simple for users to
+    combine all the above tools to run studies based in two distinct
+    modules:</p>
+    <list list-type="bullet">
+      <list-item>
+        <p>backtesting studies</p>
+      </list-item>
+      <list-item>
+        <p>live-deployments</p>
+      </list-item>
+    </list>
+    <p>Both of these modules and their respective data management
+    systems are built on top of
+    <ext-link ext-link-type="uri" xlink:href="https://www.freqtrade.io/en/latest/"><monospace>Freqtrade</monospace></ext-link>,
+    a mature and actively developed cryptocurrency trading software.
+    This means that <monospace>FreqAI</monospace> benefits from a wide
+    range of tangential/disparate feature developments such as:</p>
+    <list list-type="bullet">
+      <list-item>
+        <p>FreqUI, a graphical interface for backtesting and live
+        monitoring</p>
+      </list-item>
+      <list-item>
+        <p>telegram control</p>
+      </list-item>
+      <list-item>
+        <p>robust database handling</p>
+      </list-item>
+      <list-item>
+        <p>futures/leverage trading</p>
+      </list-item>
+      <list-item>
+        <p>dollar cost averaging</p>
+      </list-item>
+      <list-item>
+        <p>trading strategy handling</p>
+      </list-item>
+      <list-item>
+        <p>a variety of free data sources via CCXT (FTX, Binance, Kucoin
+        etc.)</p>
+      </list-item>
+    </list>
+    <p>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,
+    <monospace>FreqAI</monospace> benefits strongly from the userbase of
+    <monospace>Freqtrade</monospace>, where most
+    <monospace>FreqAI</monospace> beta-testers/developers originated.
+    This symbiotic relationship between <monospace>Freqtrade</monospace>
+    and <monospace>FreqAI</monospace> ignited a thoroughly tested
+    <ext-link ext-link-type="uri" xlink:href="https://github.com/freqtrade/freqtrade/pull/6832"><monospace>beta</monospace></ext-link>,
+    which demanded a four month beta and
+    <ext-link ext-link-type="uri" xlink:href="https://www.freqtrade.io/en/latest/freqai/">comprehensive
+    documentation</ext-link> containing:</p>
+    <list list-type="bullet">
+      <list-item>
+        <p>numerous example scripts</p>
+      </list-item>
+      <list-item>
+        <p>a full parameter table</p>
+      </list-item>
+      <list-item>
+        <p>methodological descriptions</p>
+      </list-item>
+      <list-item>
+        <p>high-resolution diagrams/figures</p>
+      </list-item>
+      <list-item>
+        <p>detailed parameter setting recommendations</p>
+      </list-item>
+    </list>
+  </sec>
+  <sec id="providing-a-reproducible-foundation-for-researchers">
+    <title>Providing a reproducible foundation for researchers</title>
+    <p><monospace>FreqAI</monospace> provides an extensible, robust,
+    framework for researchers and citizen data scientists. The
+    <monospace>FreqAI</monospace> sandbox enables rapid conception and
+    testing of exotic hypotheses. From a research perspective,
+    <monospace>FreqAI</monospace> handles the multitude of logistics
+    associated with live deployments, historical backtesting, and
+    feature engineering. With <monospace>FreqAI</monospace>, 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 <monospace>FreqAI</monospace> enables
+    reproducible scientific studies. This reproducibility component is
+    essential to general scientific advancement in time-series
+    forecasting for chaotic systems.</p>
+  </sec>
+</sec>
+<sec id="technical-details">
+  <title>Technical details</title>
+  <p>Typical users configure <monospace>FreqAI</monospace> via two
+  files:</p>
+  <list list-type="order">
+    <list-item>
+      <p>A <monospace>configuration</monospace> file
+      (<monospace>--config</monospace>) which provides access to the
+      full parameter list available
+      <ext-link ext-link-type="uri" xlink:href="https://www.freqtrade.io/en/latest/freqai/">here</ext-link>:</p>
+    </list-item>
+  </list>
+  <list list-type="bullet">
+    <list-item>
+      <p>control high-level feature engineering</p>
+    </list-item>
+    <list-item>
+      <p>customize adaptive modeling techniques</p>
+    </list-item>
+    <list-item>
+      <p>set any model training parameters available in third-party
+      libraries</p>
+    </list-item>
+    <list-item>
+      <p>manage adaptive modeling parameters (retrain frequency,
+      training window size, continual learning, etc.)</p>
+    </list-item>
+  </list>
+  <list list-type="order">
+    <list-item>
+      <label>2.</label>
+      <p>A strategy file (<monospace>--strategy</monospace>) where
+      users:</p>
+    </list-item>
+  </list>
+  <list list-type="bullet">
+    <list-item>
+      <p>list of the base training features</p>
+    </list-item>
+    <list-item>
+      <p>set standard technical-analysis strategies</p>
+    </list-item>
+    <list-item>
+      <p>control trade entry/exit criteria</p>
+    </list-item>
+  </list>
+  <p>With these two files, most users can exploit a wide range of
+  pre-existing integrations in <monospace>Catboost</monospace> and 7
+  other libraries with a simple command:</p>
+  <preformat>freqtrade trade --config config_freqai.example.json --strategy FreqaiExampleStrategy --freqaimodel CatboostRegressor</preformat>
+  <p>Advanced users will edit one of the existing
+  <monospace>--freqaimodel</monospace> files, which are simply an
+  children of the <monospace>IFreqaiModel</monospace> (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
+  <monospace>FreqAI</monospace>, which is comprised of three distinct
+  Python objects:</p>
+  <list list-type="bullet">
+    <list-item>
+      <p><monospace>IFreqaiModel</monospace></p>
+      <list list-type="bullet">
+        <list-item>
+          <p>A singular long-lived object containing all the necessary
+          logic to collect data, store data, process data, engineer
+          features, run training, and inference models.</p>
+        </list-item>
+      </list>
+    </list-item>
+    <list-item>
+      <p><monospace>FreqaiDataKitchen</monospace></p>
+      <list list-type="bullet">
+        <list-item>
+          <p>A short-lived object which is uniquely created for each
+          asset/model. Beyond metadata, it also contains a variety of
+          data processing tools.</p>
+        </list-item>
+      </list>
+    </list-item>
+    <list-item>
+      <p><monospace>FreqaiDataDrawer</monospace></p>
+      <list list-type="bullet">
+        <list-item>
+          <p>Singular long-lived object containing all the historical
+          predictions, models, and save/load methods.</p>
+        </list-item>
+      </list>
+    </list-item>
+  </list>
+  <p>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
+  <monospace>IFreqaiModel</monospace> 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 <monospace>IFreqaiModel</monospace> to help
+  them craft unique data structures and training procedures.</p>
+  <p>The class structure and algorithmic details are depicted in the
+  following diagram:</p>
+  <p><named-content content-type="image">image</named-content>
+  <italic>Class diagram summarizing object interactions in
+  FreqAI</italic></p>
+</sec>
+<sec id="online-documentation">
+  <title>Online documentation</title>
+  <p>The documentation for
+  <ext-link ext-link-type="uri" xlink:href="https://www.freqtrade.io/en/latest/freqai/"><monospace>FreqAI</monospace></ext-link>
+  is available online at
+  <ext-link ext-link-type="uri" xlink:href="https://www.freqtrade.io/en/latest/freqai/">https://www.freqtrade.io/en/latest/freqai/</ext-link>
+  and covers a wide range of materials:</p>
+  <list list-type="bullet">
+    <list-item>
+      <p>Quick-start with a single command and example files -
+      (beginners)</p>
+    </list-item>
+    <list-item>
+      <p>Introduction to the feature engineering interface and basic
+      configurations - (intermediate users)</p>
+    </list-item>
+    <list-item>
+      <p>Parameter table with indepth descriptions and default parameter
+      setting recommendations - (intermediate users)</p>
+    </list-item>
+    <list-item>
+      <p>Data analysis and post-processing - (advanced users)</p>
+    </list-item>
+    <list-item>
+      <p>Methodological considerations complemented by high resolution
+      figures - (advanced users)</p>
+    </list-item>
+    <list-item>
+      <p>Instructions for integrating third party machine learning
+      libraries into custom prediction models - (advanced users)</p>
+    </list-item>
+    <list-item>
+      <p>Software architectural description with class diagram -
+      (developers)</p>
+    </list-item>
+    <list-item>
+      <p>File structure descriptions - (developers)</p>
+    </list-item>
+  </list>
+  <p>The docs direct users to a variety of pre-made examples which
+  integrate <monospace>Catboost</monospace>,
+  <monospace>LightGBM</monospace>, <monospace>XGBoost</monospace>,
+  <monospace>Sklearn</monospace>,
+  <monospace>stable_baselines3</monospace>,
+  <monospace>torch</monospace>, <monospace>tensorflow</monospace>.
+  Meanwhile, developers will also find thorough docstrings and type
+  hinting throughout the source code to aid in code readability and
+  customization.</p>
+  <p><monospace>FreqAI</monospace> also benefits from a strong support
+  network of users and developers on the
+  <ext-link ext-link-type="uri" xlink:href="https://discord.gg/w6nDM6cM4y"><monospace>Freqtrade</monospace>
+  discord</ext-link> as well as on the
+  <ext-link ext-link-type="uri" xlink:href="https://discord.gg/xE4RMg4QYw"><monospace>FreqAI</monospace>
+  discord</ext-link>. Within the <monospace>FreqAI</monospace> discord,
+  users will find a deep and easily searched knowledge base containing
+  common errors. But more importantly, users in the
+  <monospace>FreqAI</monospace> discord share anectdotal and
+  quantitative observations which compare performance between various
+  third-party libraries and methods.</p>
+</sec>
+<sec id="state-of-the-field">
+  <title>State of the field</title>
+  <p>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
+  <ext-link ext-link-type="uri" xlink:href="https://tensortradex.readthedocs.io/en/latest/"><monospace>tensortrade</monospace></ext-link>
+  (<xref alt="Tensortrade, 2022" rid="ref-tensortrade" ref-type="bibr"><italic>Tensortrade</italic>,
+  2022</xref>) and
+  <ext-link ext-link-type="uri" xlink:href="https://github.com/AI4Finance-Foundation/FinRL"><monospace>FinRL</monospace></ext-link>
+  (<xref alt="AI4Finance-Foundation, 2022" rid="ref-finrl" ref-type="bibr"><italic>AI4Finance-Foundation</italic>,
+  2022</xref>) 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.
+  <monospace>FreqAI</monospace> fills the gap.</p>
+</sec>
+<sec id="on-going-research">
+  <title>On-going research</title>
+  <p>Emergent Methods, based in Arvada CO, is actively using
+  <monospace>FreqAI</monospace> 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 <monospace>CatboosRegressor</monospace>,
+  <monospace>LightGBMRegressor</monospace>, and
+  <monospace>XGBoostRegressor</monospace>. Results from these studies
+  are on track for publication in scientific journals as well as more
+  general data science blogs (e.g. Medium).</p>
+</sec>
+<sec id="installing-and-running-freqai">
+  <title>Installing and running <monospace>FreqAI</monospace></title>
+  <p><monospace>FreqAI</monospace> is automatically installed with
+  <monospace>Freqtrade</monospace> using the following commands on linux
+  systems:</p>
+  <preformat>git clone git@github.com:freqtrade/freqtrade.git
+cd freqtrade
+./setup.sh -i</preformat>
+  <p>However, <monospace>FreqAI</monospace> also benefits from
+  <monospace>Freqtrade</monospace> docker distributions, and can be run
+  with docker by pulling the stable or develop images from
+  <monospace>Freqtrade</monospace> distributions.</p>
+</sec>
+<sec id="funding-sources">
+  <title>Funding sources</title>
+  <p><ext-link ext-link-type="uri" xlink:href="https://www.freqtrade.io/en/latest/freqai/"><monospace>FreqAI</monospace></ext-link>
+  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.</p>
+</sec>
+<sec id="acknowledgements">
+  <title>Acknowledgements</title>
+  <p>We would like to acknowledge various beta testers of
+  <monospace>FreqAI</monospace>:</p>
+  <list list-type="bullet">
+    <list-item>
+      <p>Richárd Józsa</p>
+    </list-item>
+    <list-item>
+      <p>Juha Nykänen</p>
+    </list-item>
+    <list-item>
+      <p>Salah Lamkadem</p>
+    </list-item>
+  </list>
+  <p>As well as various <monospace>Freqtrade</monospace>
+  <ext-link ext-link-type="uri" xlink:href="https://github.com/freqtrade/freqtrade/graphs/contributors">developers</ext-link>
+  maintaining tangential, yet essential, modules.</p>
+</sec>
+</body>
+<back>
+<ref-list>
+  <ref id="ref-scikit-learn">
+    <element-citation publication-type="article-journal">
+      <person-group person-group-type="author">
+        <name><surname>Pedregosa</surname><given-names>F.</given-names></name>
+        <name><surname>Varoquaux</surname><given-names>G.</given-names></name>
+        <name><surname>Gramfort</surname><given-names>A.</given-names></name>
+        <name><surname>Michel</surname><given-names>V.</given-names></name>
+        <name><surname>Thirion</surname><given-names>B.</given-names></name>
+        <name><surname>Grisel</surname><given-names>O.</given-names></name>
+        <name><surname>Blondel</surname><given-names>M.</given-names></name>
+        <name><surname>Prettenhofer</surname><given-names>P.</given-names></name>
+        <name><surname>Weiss</surname><given-names>R.</given-names></name>
+        <name><surname>Dubourg</surname><given-names>V.</given-names></name>
+        <name><surname>Vanderplas</surname><given-names>J.</given-names></name>
+        <name><surname>Passos</surname><given-names>A.</given-names></name>
+        <name><surname>Cournapeau</surname><given-names>D.</given-names></name>
+        <name><surname>Brucher</surname><given-names>M.</given-names></name>
+        <name><surname>Perrot</surname><given-names>M.</given-names></name>
+        <name><surname>Duchesnay</surname><given-names>E.</given-names></name>
+      </person-group>
+      <article-title>Scikit-learn: Machine learning in Python</article-title>
+      <source>Journal of Machine Learning Research</source>
+      <year iso-8601-date="2011">2011</year>
+      <volume>12</volume>
+      <fpage>2825</fpage>
+      <lpage>2830</lpage>
+    </element-citation>
+  </ref>
+  <ref id="ref-catboost">
+    <element-citation publication-type="paper-conference">
+      <person-group person-group-type="author">
+        <name><surname>Prokhorenkova</surname><given-names>Liudmila</given-names></name>
+        <name><surname>Gusev</surname><given-names>Gleb</given-names></name>
+        <name><surname>Vorobev</surname><given-names>Aleksandr</given-names></name>
+        <name><surname>Dorogush</surname><given-names>Anna Veronika</given-names></name>
+        <name><surname>Gulin</surname><given-names>Andrey</given-names></name>
+      </person-group>
+      <article-title>CatBoost: Unbiased boosting with categorical features</article-title>
+      <source>Proceedings of the 32nd international conference on neural information processing systems</source>
+      <publisher-name>Curran Associates Inc.</publisher-name>
+      <publisher-loc>Red Hook, NY, USA</publisher-loc>
+      <year iso-8601-date="2018">2018</year>
+      <fpage>6639</fpage>
+      <lpage>6649</lpage>
+    </element-citation>
+  </ref>
+  <ref id="ref-lightgbm">
+    <element-citation publication-type="article-journal">
+      <person-group person-group-type="author">
+        <name><surname>Ke</surname><given-names>Guolin</given-names></name>
+        <name><surname>Meng</surname><given-names>Qi</given-names></name>
+        <name><surname>Finley</surname><given-names>Thomas</given-names></name>
+        <name><surname>Wang</surname><given-names>Taifeng</given-names></name>
+        <name><surname>Chen</surname><given-names>Wei</given-names></name>
+        <name><surname>Ma</surname><given-names>Weidong</given-names></name>
+        <name><surname>Ye</surname><given-names>Qiwei</given-names></name>
+        <name><surname>Liu</surname><given-names>Tie-Yan</given-names></name>
+      </person-group>
+      <article-title>Lightgbm: A highly efficient gradient boosting decision tree</article-title>
+      <source>Advances in neural information processing systems</source>
+      <year iso-8601-date="2017">2017</year>
+      <volume>30</volume>
+      <fpage>3146</fpage>
+      <lpage>3154</lpage>
+    </element-citation>
+  </ref>
+  <ref id="ref-xgboost">
+    <element-citation publication-type="paper-conference">
+      <person-group person-group-type="author">
+        <name><surname>Chen</surname><given-names>Tianqi</given-names></name>
+        <name><surname>Guestrin</surname><given-names>Carlos</given-names></name>
+      </person-group>
+      <article-title>XGBoost: A scalable tree boosting system</article-title>
+      <source>Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining</source>
+      <publisher-name>ACM</publisher-name>
+      <publisher-loc>New York, NY, USA</publisher-loc>
+      <year iso-8601-date="2016">2016</year>
+      <isbn>978-1-4503-4232-2</isbn>
+      <uri>http://doi.acm.org/10.1145/2939672.2939785</uri>
+      <pub-id pub-id-type="doi">10.1145/2939672.2939785</pub-id>
+      <fpage>785</fpage>
+      <lpage>794</lpage>
+    </element-citation>
+  </ref>
+  <ref id="ref-stable-baselines3">
+    <element-citation publication-type="article-journal">
+      <person-group person-group-type="author">
+        <name><surname>Raffin</surname><given-names>Antonin</given-names></name>
+        <name><surname>Hill</surname><given-names>Ashley</given-names></name>
+        <name><surname>Gleave</surname><given-names>Adam</given-names></name>
+        <name><surname>Kanervisto</surname><given-names>Anssi</given-names></name>
+        <name><surname>Ernestus</surname><given-names>Maximilian</given-names></name>
+        <name><surname>Dormann</surname><given-names>Noah</given-names></name>
+      </person-group>
+      <article-title>Stable-Baselines3: Reliable reinforcement learning implementations</article-title>
+      <source>Journal of Machine Learning Research</source>
+      <year iso-8601-date="2021">2021</year>
+      <volume>22</volume>
+      <issue>268</issue>
+      <uri>http://jmlr.org/papers/v22/20-1364.html</uri>
+      <fpage>1</fpage>
+      <lpage>8</lpage>
+    </element-citation>
+  </ref>
+  <ref id="ref-openai">
+    <element-citation>
+      <person-group person-group-type="author">
+        <name><surname>Brockman</surname><given-names>Greg</given-names></name>
+        <name><surname>Cheung</surname><given-names>Vicki</given-names></name>
+        <name><surname>Pettersson</surname><given-names>Ludwig</given-names></name>
+        <name><surname>Schneider</surname><given-names>Jonas</given-names></name>
+        <name><surname>Schulman</surname><given-names>John</given-names></name>
+        <name><surname>Tang</surname><given-names>Jie</given-names></name>
+        <name><surname>Zaremba</surname><given-names>Wojciech</given-names></name>
+      </person-group>
+      <article-title>OpenAI gym</article-title>
+      <year iso-8601-date="2016">2016</year>
+      <uri>https://arxiv.org/abs/1606.01540</uri>
+    </element-citation>
+  </ref>
+  <ref id="ref-tensorflow">
+    <element-citation>
+      <person-group person-group-type="author">
+        <name><surname>Abadi</surname><given-names>Martín</given-names></name>
+        <name><surname>Agarwal</surname><given-names>Ashish</given-names></name>
+        <name><surname>Barham</surname><given-names>Paul</given-names></name>
+        <name><surname>Brevdo</surname><given-names>Eugene</given-names></name>
+        <name><surname>Chen</surname><given-names>Zhifeng</given-names></name>
+        <name><surname>Citro</surname><given-names>Craig</given-names></name>
+        <name><surname>Corrado</surname><given-names>Greg S.</given-names></name>
+        <name><surname>Davis</surname><given-names>Andy</given-names></name>
+        <name><surname>Dean</surname><given-names>Jeffrey</given-names></name>
+        <name><surname>Devin</surname><given-names>Matthieu</given-names></name>
+        <name><surname>Ghemawat</surname><given-names>Sanjay</given-names></name>
+        <name><surname>Goodfellow</surname><given-names>Ian</given-names></name>
+        <name><surname>Harp</surname><given-names>Andrew</given-names></name>
+        <name><surname>Irving</surname><given-names>Geoffrey</given-names></name>
+        <name><surname>Isard</surname><given-names>Michael</given-names></name>
+        <name><surname>Jia</surname><given-names>Yangqing</given-names></name>
+        <name><surname>Jozefowicz</surname><given-names>Rafal</given-names></name>
+        <name><surname>Kaiser</surname><given-names>Lukasz</given-names></name>
+        <name><surname>Kudlur</surname><given-names>Manjunath</given-names></name>
+        <name><surname>Levenberg</surname><given-names>Josh</given-names></name>
+        <name><surname>Mané</surname><given-names>Dandelion</given-names></name>
+        <name><surname>Monga</surname><given-names>Rajat</given-names></name>
+        <name><surname>Moore</surname><given-names>Sherry</given-names></name>
+        <name><surname>Murray</surname><given-names>Derek</given-names></name>
+        <name><surname>Olah</surname><given-names>Chris</given-names></name>
+        <name><surname>Schuster</surname><given-names>Mike</given-names></name>
+        <name><surname>Shlens</surname><given-names>Jonathon</given-names></name>
+        <name><surname>Steiner</surname><given-names>Benoit</given-names></name>
+        <name><surname>Sutskever</surname><given-names>Ilya</given-names></name>
+        <name><surname>Talwar</surname><given-names>Kunal</given-names></name>
+        <name><surname>Tucker</surname><given-names>Paul</given-names></name>
+        <name><surname>Vanhoucke</surname><given-names>Vincent</given-names></name>
+        <name><surname>Vasudevan</surname><given-names>Vijay</given-names></name>
+        <name><surname>Viégas</surname><given-names>Fernanda</given-names></name>
+        <name><surname>Vinyals</surname><given-names>Oriol</given-names></name>
+        <name><surname>Warden</surname><given-names>Pete</given-names></name>
+        <name><surname>Wattenberg</surname><given-names>Martin</given-names></name>
+        <name><surname>Wicke</surname><given-names>Martin</given-names></name>
+        <name><surname>Yu</surname><given-names>Yuan</given-names></name>
+        <name><surname>Zheng</surname><given-names>Xiaoqiang</given-names></name>
+      </person-group>
+      <article-title>TensorFlow: Large-scale machine learning on heterogeneous systems</article-title>
+      <year iso-8601-date="2015">2015</year>
+      <uri>https://www.tensorflow.org/</uri>
+    </element-citation>
+  </ref>
+  <ref id="ref-pytorch">
+    <element-citation publication-type="chapter">
+      <person-group person-group-type="author">
+        <name><surname>Paszke</surname><given-names>Adam</given-names></name>
+        <name><surname>Gross</surname><given-names>Sam</given-names></name>
+        <name><surname>Massa</surname><given-names>Francisco</given-names></name>
+        <name><surname>Lerer</surname><given-names>Adam</given-names></name>
+        <name><surname>Bradbury</surname><given-names>James</given-names></name>
+        <name><surname>Chanan</surname><given-names>Gregory</given-names></name>
+        <name><surname>Killeen</surname><given-names>Trevor</given-names></name>
+        <name><surname>Lin</surname><given-names>Zeming</given-names></name>
+        <name><surname>Gimelshein</surname><given-names>Natalia</given-names></name>
+        <name><surname>Antiga</surname><given-names>Luca</given-names></name>
+        <name><surname>Desmaison</surname><given-names>Alban</given-names></name>
+        <name><surname>Kopf</surname><given-names>Andreas</given-names></name>
+        <name><surname>Yang</surname><given-names>Edward</given-names></name>
+        <name><surname>DeVito</surname><given-names>Zachary</given-names></name>
+        <name><surname>Raison</surname><given-names>Martin</given-names></name>
+        <name><surname>Tejani</surname><given-names>Alykhan</given-names></name>
+        <name><surname>Chilamkurthy</surname><given-names>Sasank</given-names></name>
+        <name><surname>Steiner</surname><given-names>Benoit</given-names></name>
+        <name><surname>Fang</surname><given-names>Lu</given-names></name>
+        <name><surname>Bai</surname><given-names>Junjie</given-names></name>
+        <name><surname>Chintala</surname><given-names>Soumith</given-names></name>
+      </person-group>
+      <article-title>PyTorch: An imperative style, high-performance deep learning library</article-title>
+      <source>Advances in neural information processing systems 32</source>
+      <person-group person-group-type="editor">
+        <name><surname>Wallach</surname><given-names>H.</given-names></name>
+        <name><surname>Larochelle</surname><given-names>H.</given-names></name>
+        <name><surname>Beygelzimer</surname><given-names>A.</given-names></name>
+        <name><surname>dAlché-Buc</surname><given-names>F.</given-names></name>
+        <name><surname>Fox</surname><given-names>E.</given-names></name>
+        <name><surname>Garnett</surname><given-names>R.</given-names></name>
+      </person-group>
+      <publisher-name>Curran Associates, Inc.</publisher-name>
+      <year iso-8601-date="2019">2019</year>
+      <uri>http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf</uri>
+      <fpage>8024</fpage>
+      <lpage>8035</lpage>
+    </element-citation>
+  </ref>
+  <ref id="ref-scipy">
+    <element-citation publication-type="article-journal">
+      <person-group person-group-type="author">
+        <name><surname>Virtanen</surname><given-names>Pauli</given-names></name>
+        <name><surname>Gommers</surname><given-names>Ralf</given-names></name>
+        <name><surname>Oliphant</surname><given-names>Travis E.</given-names></name>
+        <name><surname>Haberland</surname><given-names>Matt</given-names></name>
+        <name><surname>Reddy</surname><given-names>Tyler</given-names></name>
+        <name><surname>Cournapeau</surname><given-names>David</given-names></name>
+        <name><surname>Burovski</surname><given-names>Evgeni</given-names></name>
+        <name><surname>Peterson</surname><given-names>Pearu</given-names></name>
+        <name><surname>Weckesser</surname><given-names>Warren</given-names></name>
+        <name><surname>Bright</surname><given-names>Jonathan</given-names></name>
+        <name><surname>van der Walt</surname><given-names>Stéfan J.</given-names></name>
+        <name><surname>Brett</surname><given-names>Matthew</given-names></name>
+        <name><surname>Wilson</surname><given-names>Joshua</given-names></name>
+        <name><surname>Millman</surname><given-names>K. Jarrod</given-names></name>
+        <name><surname>Mayorov</surname><given-names>Nikolay</given-names></name>
+        <name><surname>Nelson</surname><given-names>Andrew R. J.</given-names></name>
+        <name><surname>Jones</surname><given-names>Eric</given-names></name>
+        <name><surname>Kern</surname><given-names>Robert</given-names></name>
+        <name><surname>Larson</surname><given-names>Eric</given-names></name>
+        <name><surname>Carey</surname><given-names>C J</given-names></name>
+        <name><surname>Polat</surname><given-names>İlhan</given-names></name>
+        <name><surname>Feng</surname><given-names>Yu</given-names></name>
+        <name><surname>Moore</surname><given-names>Eric W.</given-names></name>
+        <name><surname>VanderPlas</surname><given-names>Jake</given-names></name>
+        <name><surname>Laxalde</surname><given-names>Denis</given-names></name>
+        <name><surname>Perktold</surname><given-names>Josef</given-names></name>
+        <name><surname>Cimrman</surname><given-names>Robert</given-names></name>
+        <name><surname>Henriksen</surname><given-names>Ian</given-names></name>
+        <name><surname>Quintero</surname><given-names>E. A.</given-names></name>
+        <name><surname>Harris</surname><given-names>Charles R.</given-names></name>
+        <name><surname>Archibald</surname><given-names>Anne M.</given-names></name>
+        <name><surname>Ribeiro</surname><given-names>Antônio H.</given-names></name>
+        <name><surname>Pedregosa</surname><given-names>Fabian</given-names></name>
+        <name><surname>van Mulbregt</surname><given-names>Paul</given-names></name>
+        <string-name>SciPy 1.0 Contributors</string-name>
+      </person-group>
+      <article-title>SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python</article-title>
+      <source>Nature Methods</source>
+      <year iso-8601-date="2020">2020</year>
+      <volume>17</volume>
+      <pub-id pub-id-type="doi">10.1038/s41592-019-0686-2</pub-id>
+      <fpage>261</fpage>
+      <lpage>272</lpage>
+    </element-citation>
+  </ref>
+  <ref id="ref-numpy">
+    <element-citation publication-type="article-journal">
+      <person-group person-group-type="author">
+        <name><surname>Harris</surname><given-names>Charles R.</given-names></name>
+        <name><surname>Millman</surname><given-names>K. Jarrod</given-names></name>
+        <name><surname>Walt</surname><given-names>Stéfan J. van der</given-names></name>
+        <name><surname>Gommers</surname><given-names>Ralf</given-names></name>
+        <name><surname>Virtanen</surname><given-names>Pauli</given-names></name>
+        <name><surname>Cournapeau</surname><given-names>David</given-names></name>
+        <name><surname>Wieser</surname><given-names>Eric</given-names></name>
+        <name><surname>Taylor</surname><given-names>Julian</given-names></name>
+        <name><surname>Berg</surname><given-names>Sebastian</given-names></name>
+        <name><surname>Smith</surname><given-names>Nathaniel J.</given-names></name>
+        <name><surname>Kern</surname><given-names>Robert</given-names></name>
+        <name><surname>Picus</surname><given-names>Matti</given-names></name>
+        <name><surname>Hoyer</surname><given-names>Stephan</given-names></name>
+        <name><surname>Kerkwijk</surname><given-names>Marten H. van</given-names></name>
+        <name><surname>Brett</surname><given-names>Matthew</given-names></name>
+        <name><surname>Haldane</surname><given-names>Allan</given-names></name>
+        <name><surname>Río</surname><given-names>Jaime Fernández del</given-names></name>
+        <name><surname>Wiebe</surname><given-names>Mark</given-names></name>
+        <name><surname>Peterson</surname><given-names>Pearu</given-names></name>
+        <name><surname>Gérard-Marchant</surname><given-names>Pierre</given-names></name>
+        <name><surname>Sheppard</surname><given-names>Kevin</given-names></name>
+        <name><surname>Reddy</surname><given-names>Tyler</given-names></name>
+        <name><surname>Weckesser</surname><given-names>Warren</given-names></name>
+        <name><surname>Abbasi</surname><given-names>Hameer</given-names></name>
+        <name><surname>Gohlke</surname><given-names>Christoph</given-names></name>
+        <name><surname>Oliphant</surname><given-names>Travis E.</given-names></name>
+      </person-group>
+      <article-title>Array programming with NumPy</article-title>
+      <source>Nature</source>
+      <publisher-name>Springer Science; Business Media LLC</publisher-name>
+      <year iso-8601-date="2020-09">2020</year><month>09</month>
+      <volume>585</volume>
+      <issue>7825</issue>
+      <uri>https://doi.org/10.1038/s41586-020-2649-2</uri>
+      <pub-id pub-id-type="doi">10.1038/s41586-020-2649-2</pub-id>
+      <fpage>357</fpage>
+      <lpage>362</lpage>
+    </element-citation>
+  </ref>
+  <ref id="ref-pandas">
+    <element-citation publication-type="paper-conference">
+      <person-group person-group-type="author">
+        <name><surname>McKinney</surname><given-names>Wes</given-names></name>
+        <name><surname>others</surname></name>
+      </person-group>
+      <article-title>Data structures for statistical computing in python</article-title>
+      <source>Proceedings of the 9th python in science conference</source>
+      <publisher-name>Austin, TX</publisher-name>
+      <year iso-8601-date="2010">2010</year>
+      <volume>445</volume>
+      <fpage>51</fpage>
+      <lpage>56</lpage>
+    </element-citation>
+  </ref>
+  <ref id="ref-finrl">
+    <element-citation publication-type="webpage">
+      <article-title>AI4Finance-foundation</article-title>
+      <year iso-8601-date="2022">2022</year>
+      <date-in-citation content-type="access-date"><year iso-8601-date="2022-09-30">2022</year><month>09</month><day>30</day></date-in-citation>
+      <uri>https://github.com/AI4Finance-Foundation/FinRL</uri>
+    </element-citation>
+  </ref>
+  <ref id="ref-tensortrade">
+    <element-citation publication-type="webpage">
+      <article-title>Tensortrade</article-title>
+      <year iso-8601-date="2022">2022</year>
+      <date-in-citation content-type="access-date"><year iso-8601-date="2022-09-30">2022</year><month>09</month><day>30</day></date-in-citation>
+      <uri>https://tensortradex.readthedocs.io/en/latest/L</uri>
+    </element-citation>
+  </ref>
+</ref-list>
+</back>
+</article>

docs/JOSS_paper/paper.md

@@ -0,0 +1,212 @@
+---
+title: '`FreqAI`: generalizing adaptive modeling for chaotic time-series market forecasts'
+tags:
+  - Python
+  - Machine Learning
+  - adaptive modeling
+  - chaotic systems
+  - time-series forecasting
+authors:
+  - name: Robert A. Caulk
+    orcid: 0000-0001-5618-8629
+    affiliation: 1, 2 
+  - name: Elin Törnquist
+    orcid: 0000-0003-3289-8604
+    affiliation: 1, 2
+  - name: Matthias Voppichler 
+    orcid: 
+    affiliation: 2
+  - name: Andrew R. Lawless
+    orcid: 
+    affiliation: 2
+  - name: Ryan McMullan
+    orcid: 
+    affiliation: 2
+  - name: Wagner Costa Santos
+    orcid: 
+    affiliation: 1, 2
+  - name: Timothy C. Pogue
+    orcid: 
+    affiliation: 1, 2
+  - name: Johan van der Vlugt
+    orcid: 
+    affiliation: 2
+  - name: Stefan P. Gehring
+    orcid: 
+    affiliation: 2
+  - name: Pascal Schmidt
+    orcid: 0000-0001-9328-4345
+    affiliation: 2
+
+<!-- affiliation: "1, 2" # (Multiple affiliations must be quoted) -->
+affiliations:
+ - name: Emergent Methods LLC, Arvada Colorado, 80005, USA
+   index: 1
+ - name: Freqtrade open source project
+   index: 2
+date: October 2022
+bibliography: paper.bib
+
+
+---
+
+
+# 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 "citizen 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`](https://www.freqtrade.io/en/latest/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`](https://www.freqtrade.io/en/latest/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 via a variety of open-sourced [exchange APIs](https://docs.ccxt.com/en/latest/manual.html#exchange-structure). Luckily, an existing open-source software, [`Freqtrade`](https://www.freqtrade.io/en/stable/), 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](assets/freqai_algo.jpg)
+*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 [@scikit-learn], Catboost [@catboost], LightGBM [@lightgbm], XGBoost [@xgboost], stable_baselines3 [@stable-baselines3], openai gym [@openai], tensorflow [@tensorflow], pytorch [@pytorch], Scipy [@scipy], Numpy [@numpy], and pandas [@pandas]. 
+
+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 details [here](https://www.freqtrade.io/en/stable/freqai/#general-approach) and [here](https://www.freqtrade.io/en/stable/freqai-developers/#project-architecture)
+* rapid and self-monitored feature engineering tools, details [here](https://www.freqtrade.io/en/stable/freqai-feature-engineering/#feature-engineering)
+* unique model features/indicators, such as the [inlier metric](https://www.freqtrade.io/en/stable/freqai-feature-engineering/#inlier-metric)
+* optimized data collection/storage algorithms, all code shown [here](https://github.com/freqtrade/freqtrade/blob/develop/freqtrade/freqai/data_drawer.py)
+* safely integrated outlier detection methods, details [here](https://www.freqtrade.io/en/stable/freqai-feature-engineering/#outlier-detection)
+* websocket communicated forecasts, details [here](https://www.freqtrade.io/en/stable/producer-consumer/)
+
+Of particular interest for researchers, `FreqAI` provides the option of large scale experimentation via an optimized [websocket communications interface](https://www.freqtrade.io/en/stable/producer-consumer/).
+
+## 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`](https://www.freqtrade.io/en/latest/), 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](https://docs.ccxt.com/en/latest/manual.html#exchange-structure) (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`](https://github.com/freqtrade/freqtrade/pull/6832), which demanded a four month beta and [comprehensive documentation](https://www.freqtrade.io/en/latest/freqai/) 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:
+
+1. A `configuration` file (`--config`) which provides access to the full parameter list available [here](https://www.freqtrade.io/en/latest/freqai/):
+  * 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.)
+
+2. 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](assets/freqai_algorithm-diagram.jpg)
+*Class diagram summarizing object interactions in FreqAI*
+
+# Online documentation
+
+The documentation for [`FreqAI`](https://www.freqtrade.io/en/latest/freqai/) is available online at [https://www.freqtrade.io/en/latest/freqai/](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](https://discord.gg/w6nDM6cM4y) as well as on the [`FreqAI` discord](https://discord.gg/xE4RMg4QYw). 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`](https://tensortradex.readthedocs.io/en/latest/) [@tensortrade] and [`FinRL`](https://github.com/AI4Finance-Foundation/FinRL) [@finrl] 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 `CatboostRegressor`, `LightGBMRegressor`, and `XGBoostRegressor`. Results from these studies are planned for submission to scientific journals as well as more general data science blogs (e.g. Medium).
+
+# Installing and running `FreqAI`
+
+`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`](https://www.freqtrade.io/en/latest/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`:
+
+- Longlong Yu (lolongcovas)
+- Richárd Józsa (richardjozsa)
+- Juha Nykänen (suikula)
+- Emre Suzen (aemr3)
+- Salah Lamkadem (ikonx)
+
+As well as various `Freqtrade` [developers](https://github.com/freqtrade/freqtrade/graphs/contributors) maintaining tangential, yet essential, modules.
+
+# References

docs/freqai-parameter-table.md

@@ -58,4 +58,4 @@ Mandatory parameters are marked as **Required** and have to be set in one of the
 | `model_reward_parameters` | Parameters used inside the user customizable `calculate_reward()` function in `ReinforcementLearner.py` <br> **Datatype:** int.
 |  |  **Extraneous parameters**
 | `keras` | If the selected model makes use of Keras (typical for Tensorflow-based prediction models), this flag needs to be activated so that the model save/loading follows Keras standards. <br> **Datatype:** Boolean. <br> Default: `False`.
-| `conv_width` | The width of a convolutional neural network input tensor. This replaces the need for shifting candles (`include_shifted_candles`) by feeding in historical data points as the second dimension of the tensor. Technically, this parameter can also be used for regressors, but it only adds computational overhead and does not change the model training/prediction. <br> **Datatype:** Integer. <br> Default: `2`.
+| `conv_width` | The width of a convolutional neural network input tensor. This replaces the need for shifting candles (`include_shifted_candles`) by feeding in historical data points as the second dimension of the tensor. Technically, this parameter can also be used for regressors, but it only adds computational overhead and does not change the model training/prediction. <br> **Datatype:** Integer. <br> Default: `2`.

freqtrade/freqai/base_models/BaseTensorFlowModel.py

@@ -3,10 +3,10 @@ from time import time
 from typing import Any
 
 from pandas import DataFrame
-
+import numpy as np
 from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
 from freqtrade.freqai.freqai_interface import IFreqaiModel
-
+import tensorflow as tf
 
 logger = logging.getLogger(__name__)
 
@@ -17,6 +17,13 @@ class BaseTensorFlowModel(IFreqaiModel):
     User *must* inherit from this class and set fit() and predict().
     """
 
+    def __init__(self, **kwargs):
+        super().__init__(config=kwargs['config'])
+        self.keras = True
+        if self.ft_params.get("DI_threshold", 0):
+            self.ft_params["DI_threshold"] = 0
+            logger.warning("DI threshold is not configured for Keras models yet. Deactivating.")
+
     def train(
         self, unfiltered_df: DataFrame, pair: str, dk: FreqaiDataKitchen, **kwargs
     ) -> Any:
@@ -68,3 +75,76 @@ class BaseTensorFlowModel(IFreqaiModel):
                     f"({end_time - start_time:.2f} secs) --------------------")
 
         return model
+
+
+class WindowGenerator:
+    def __init__(
+        self,
+        input_width,
+        label_width,
+        shift,
+        train_df=None,
+        val_df=None,
+        test_df=None,
+        train_labels=None,
+        val_labels=None,
+        test_labels=None,
+        batch_size=None,
+    ):
+        # Store the raw data.
+        self.train_df = train_df
+        self.val_df = val_df
+        self.test_df = test_df
+        self.train_labels = train_labels
+        self.val_labels = val_labels
+        self.test_labels = test_labels
+        self.batch_size = batch_size
+        self.input_width = input_width
+        self.label_width = label_width
+        self.shift = shift
+        self.total_window_size = input_width + shift
+        self.input_slice = slice(0, input_width)
+        self.input_indices = np.arange(self.total_window_size)[self.input_slice]
+
+    def make_dataset(self, data, labels=None):
+        data = np.array(data, dtype=np.float32)
+        if labels is not None:
+            labels = np.array(labels, dtype=np.float32)
+        ds = tf.keras.preprocessing.timeseries_dataset_from_array(
+            data=data,
+            targets=labels,
+            sequence_length=self.total_window_size,
+            sequence_stride=1,
+            sampling_rate=1,
+            shuffle=False,
+            batch_size=self.batch_size,
+        )
+
+        return ds
+
+    @property
+    def train(self):
+        return self.make_dataset(self.train_df, self.train_labels)
+
+    @property
+    def val(self):
+        return self.make_dataset(self.val_df, self.val_labels)
+
+    @property
+    def test(self):
+        return self.make_dataset(self.test_df, self.test_labels)
+
+    @property
+    def inference(self):
+        return self.make_dataset(self.test_df)
+
+    @property
+    def example(self):
+        """Get and cache an example batch of `inputs, labels` for plotting."""
+        result = getattr(self, "_example", None)
+        if result is None:
+            # No example batch was found, so get one from the `.train` dataset
+            result = next(iter(self.train))
+            # And cache it for next time
+            self._example = result
+        return result

freqtrade/freqai/data_kitchen.py

@@ -77,9 +77,10 @@ class FreqaiDataKitchen:
         self.backtest_predictions_folder: str = "backtesting_predictions"
         self.live = live
         self.pair = pair
+        self.model_save_type = self.freqai_config.get('model_save_type', 'joblib')
 
         self.svm_model: linear_model.SGDOneClassSVM = None
-        self.keras: bool = self.freqai_config.get("keras", False)
+        # self.model_save_type: bool = self.freqai_config.get("keras", False)
         self.set_all_pairs()
         if not self.live:
             if not self.config["timerange"]:
@@ -569,7 +570,7 @@ class FreqaiDataKitchen:
         predict: bool = If true, inference an existing SVM model, else construct one
         """
 
-        if self.keras:
+        if self.model_save_type == 'keras':
             logger.warning(
                 "SVM outlier removal not currently supported for Keras based models. "
                 "Skipping user requested function."

freqtrade/freqai/freqai_interface.py

@@ -73,10 +73,10 @@ class IFreqaiModel(ABC):
         self.identifier: str = self.freqai_info.get("identifier", "no_id_provided")
         self.scanning = False
         self.ft_params = self.freqai_info["feature_parameters"]
-        self.keras: bool = self.freqai_info.get("keras", False)
+        # self.keras: bool = self.freqai_info.get("keras", False)
-        if self.keras and self.ft_params.get("DI_threshold", 0):
+        # if self.keras and self.ft_params.get("DI_threshold", 0):
-            self.ft_params["DI_threshold"] = 0
+        #     self.ft_params["DI_threshold"] = 0
-            logger.warning("DI threshold is not configured for Keras models yet. Deactivating.")
+        #     logger.warning("DI threshold is not configured for Keras models yet. Deactivating.")
         self.CONV_WIDTH = self.freqai_info.get("conv_width", 2)
         if self.ft_params.get("inlier_metric_window", 0):
             self.CONV_WIDTH = self.ft_params.get("inlier_metric_window", 0) * 2
@@ -645,7 +645,8 @@ class IFreqaiModel(ABC):
 
         # # for keras type models, the conv_window needs to be prepended so
         # # viewing is correct in frequi
-        if self.freqai_info.get('keras', False) or self.ft_params.get('inlier_metric_window', 0):
+        if (not self.freqai_info.get('model_save_type', 'joblib') or
+                self.ft_params.get('inlier_metric_window', 0)):
             n_lost_points = self.freqai_info.get('conv_width', 2)
             zeros_df = DataFrame(np.zeros((n_lost_points, len(hist_preds_df.columns))),
                                  columns=hist_preds_df.columns)

freqtrade/freqai/prediction_models/CNNPredictionModel.py

@@ -0,0 +1,144 @@
+import logging
+from typing import Any, Dict, Tuple
+
+from pandas import DataFrame
+from freqtrade.exceptions import OperationalException
+from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
+import tensorflow as tf
+from freqtrade.freqai.base_models.BaseTensorFlowModel import BaseTensorFlowModel, WindowGenerator
+from tensorflow.keras.layers import Input, Conv1D, Dense
+from tensorflow.keras.models import Model
+import numpy as np
+
+logger = logging.getLogger(__name__)
+
+# tf.config.run_functions_eagerly(True)
+# tf.data.experimental.enable_debug_mode()
+
+MAX_EPOCHS = 10
+
+
+class CNNPredictionModel(BaseTensorFlowModel):
+    """
+    User created prediction model. The class needs to override three necessary
+    functions, predict(), fit().
+    """
+
+    def fit(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen) -> Any:
+        """
+        User sets up the training and test data to fit their desired model here
+        :params:
+        :data_dictionary: the dictionary constructed by DataHandler to hold
+        all the training and test data/labels.
+        """
+        train_df = data_dictionary["train_features"]
+        train_labels = data_dictionary["train_labels"]
+        test_df = data_dictionary["test_features"]
+        test_labels = data_dictionary["test_labels"]
+        n_labels = len(train_labels.columns)
+        if n_labels > 1:
+            raise OperationalException(
+                "Neural Net not yet configured for multi-targets. Please "
+                " reduce number of targets to 1 in strategy."
+            )
+
+        n_features = len(data_dictionary["train_features"].columns)
+        BATCH_SIZE = self.freqai_info.get("batch_size", 64)
+        input_dims = [BATCH_SIZE, self.CONV_WIDTH, n_features]
+
+        w1 = WindowGenerator(
+            input_width=self.CONV_WIDTH,
+            label_width=1,
+            shift=1,
+            train_df=train_df,
+            val_df=test_df,
+            train_labels=train_labels,
+            val_labels=test_labels,
+            batch_size=BATCH_SIZE,
+        )
+
+        model = self.create_model(input_dims, n_labels)
+
+        steps_per_epoch = np.ceil(len(test_df) / BATCH_SIZE)
+        lr_schedule = tf.keras.optimizers.schedules.InverseTimeDecay(
+            0.001, decay_steps=steps_per_epoch * 1000, decay_rate=1, staircase=False
+        )
+
+        early_stopping = tf.keras.callbacks.EarlyStopping(
+            monitor="loss", patience=3, mode="min", min_delta=0.0001
+        )
+
+        model.compile(
+            loss=tf.losses.MeanSquaredError(),
+            optimizer=tf.optimizers.Adam(lr_schedule),
+            metrics=[tf.metrics.MeanAbsoluteError()],
+        )
+
+        model.fit(
+            w1.train,
+            epochs=MAX_EPOCHS,
+            shuffle=False,
+            validation_data=w1.val,
+            callbacks=[early_stopping],
+            verbose=1,
+        )
+
+        return model
+
+    def predict(
+        self, unfiltered_dataframe: DataFrame, dk: FreqaiDataKitchen, first=True
+    ) -> Tuple[DataFrame, DataFrame]:
+        """
+        Filter the prediction features data and predict with it.
+        :param: unfiltered_dataframe: Full dataframe for the current backtest period.
+        :return:
+        :predictions: np.array of predictions
+        :do_predict: np.array of 1s and 0s to indicate places where freqai needed to remove
+        data (NaNs) or felt uncertain about data (PCA and DI index)
+        """
+
+        dk.find_features(unfiltered_dataframe)
+        filtered_dataframe, _ = dk.filter_features(
+            unfiltered_dataframe, dk.training_features_list, training_filter=False
+        )
+        filtered_dataframe = dk.normalize_data_from_metadata(filtered_dataframe)
+        dk.data_dictionary["prediction_features"] = filtered_dataframe
+
+        # optional additional data cleaning/analysis
+        self.data_cleaning_predict(dk, filtered_dataframe)
+
+        if first:
+            full_df = dk.data_dictionary["prediction_features"]
+
+            w1 = WindowGenerator(
+                input_width=self.CONV_WIDTH,
+                label_width=1,
+                shift=1,
+                test_df=full_df,
+                batch_size=len(full_df),
+            )
+
+            predictions = self.model.predict(w1.inference)
+            len_diff = len(dk.do_predict) - len(predictions)
+            if len_diff > 0:
+                dk.do_predict = dk.do_predict[len_diff:]
+
+        else:
+            data = dk.data_dictionary["prediction_features"]
+            data = tf.expand_dims(data, axis=0)
+            predictions = self.model(data, training=False)
+
+        predictions = predictions[:, 0, 0]
+        pred_df = DataFrame(predictions, columns=dk.label_list)
+
+        pred_df = dk.denormalize_labels_from_metadata(pred_df)
+
+        return (pred_df, np.ones(len(pred_df)))
+
+    def create_model(self, input_dims, n_labels) -> Any:
+
+        input_layer = Input(shape=(input_dims[1], input_dims[2]))
+        Layer_1 = Conv1D(filters=32, kernel_size=(self.CONV_WIDTH,), activation="relu")(input_layer)
+        Layer_3 = Dense(units=32, activation="relu")(Layer_1)
+        output_layer = Dense(units=n_labels)(Layer_3)
+        return Model(inputs=input_layer, outputs=output_layer)

mkdocs.yml

@@ -31,6 +31,7 @@ nav:
         - Running FreqAI: freqai-running.md
         - Reinforcement Learning: freqai-reinforcement-learning.md
         - Developer guide: freqai-developers.md
+        - JOSS paper: paper.md
     - Short / Leverage: leverage.md
     - Utility Sub-commands: utils.md
     - Plotting: plotting.md