Awesome Open Source
Awesome Open Source
...Minimizing the mean square error on future experience.  - Richard S. Sutton


Scalable event-driven RL-friendly backtesting library. Build on top of Backtrader with OpenAI Gym environment API.

Backtrader is open-source algorithmic trading library:
Documentation and community:

OpenAI Gym is..., well, everyone knows Gym:
Documentation and community:


General purpose of this project is to provide gym-integrated framework for running reinforcement learning experiments in [close to] real world algorithmic trading environments.

Code presented here is research/development grade.
Can be unstable, buggy, poor performing and is subject to change.

Note that this package is neither out-of-the-box-moneymaker, nor it provides ready-to-converge RL solutions.
Think of it as framework for setting experiments with complex non-stationary stochastic environments.

As a research project BTGym in its current stage can hardly deliver easy end-user experience in as sense that
setting meaninfull  experiments will require some practical programming experience as well as general knowledge
of reinforcement learning theory.

News and update notes



It is highly recommended to run BTGym in designated virtual environment.

Clone or copy btgym repository to local disk, cd to it and run: pip install -e . to install package and all dependencies:

git clone

cd btgym

pip install -e .

To update to latest version::

cd btgym

git pull

pip install --upgrade -e .
  1. BTGym requres Matplotlib version 2.0.2, downgrade your installation if you have version 2.1:

    pip install matplotlib==2.0.2

  2. LSOF utility should be installed to your OS, which can not be the default case for some Linux distributives, see:


Making gym environment with all parmeters set to defaults is as simple as:

from btgym import BTgymEnv

MyEnvironment = BTgymEnv(filename='../examples/data/DAT_ASCII_EURUSD_M1_2016.csv',)

Adding more controls may look like:

from gym import spaces
from btgym import BTgymEnv

MyEnvironment = BTgymEnv(filename='../examples/data/DAT_ASCII_EURUSD_M1_2016.csv',
                         episode_duration={'days': 2, 'hours': 23, 'minutes': 55},
See more options at Documentation: Quickstart >>
and how-to's in Examples directory >>.

General description

Problem setting

  • Discrete actions setup: consider setup with one riskless asset acting as broker account cash and K (by default - one) risky assets. For every risky asset there exists track of historic price records referred as data-line. Apart from assets data lines there [optionally] exists number of exogenous data lines holding some information and statistics, e.g. economic indexes, encoded news, macroeconomic indicators, weather forecasts etc. which are considered relevant to decision-making. It is supposed for this setup that:

    1. there is no interest rates for any asset;
    2. broker actions are fixed-size market orders (buy, sell, close); short selling is permitted;
    3. transaction costs are modelled via broker commission;
    4. 'market liquidity' and 'capital impact' assumptions are met;
    5. time indexes match for all data lines provided;
  • The problem is modelled as discrete-time finite-horizon partially observable Markov decision process for equity/currency trading:

    • for every asset traded agent action space is discrete (0:hold[do nothing], 1:buy, 2:sell, 3:close[position]);
    • environment is episodic: maximum episode duration and episode termination conditions are set;
    • for every timestep of the episode agent is given environment state observation as tensor of last m time-embedded preprocessed values for every data-line included and emits actions according some stochastic policy.
    • agent's goal is to maximize expected cumulative capital by learning optimal policy;
  • Continuous actions setup[BETA]: this setup closely relates to continuous portfolio optimisation problem definition; it differs from setup above in:

    1. base broker actions are real numbers: a[i] in [0,1], 0<=i<=K, SUM{a[i]} = 1 for K risky assets added; each action is a market target order to adjust portfolio to get share a[i]*100% for i-th asset;
    2. entire single-step broker action is dictionary of form: {cash_name: a[0], asset_name_1: a[1], ..., asset_name_K: a[K]};
    3. short selling is not permitted;
  • For RL it implies having continuous action space as K+1 dim vector.

Data selection options for backtest agent training:

Notice: data shaping approach is under development, expect some changes. [7.01.18]

  • random sampling: historic price change dataset is divided to training, cross-validation and testing subsets. Since agent actions do not influence market, it is possible to randomly sample continuous subset of training data for every episode. [Seems to be] most data-efficient method. Cross-validation and testing performed later as usual on most "recent" data;
  • sequential sampling: full dataset is feeded sequentially as if agent is performing real-time trading, episode by episode. Most reality-like, least data-efficient, natural non-stationarity remedy.
  • sliding time-window sampling: mixture of above, episde is sampled randomly from comparatively short time period, sliding from furthest to most recent training data. Should be less prone to overfitting than random sampling.

Documentation and Community

Known bugs and limitations:

  • requres Matplotlib version 2.0.2;
  • matplotlib backend warning: appears when importing pyplot and using %matplotlib inline magic before btgym import. It's recommended to import btacktrader and btgym first to ensure proper backend choice;
  • not tested with Python < 3.5;
  • doesn't seem to work correctly under Windows; partially done
  • by default, is configured to accept Forex 1 min. data from;
  • only random data sampling is implemented;
  • no built-in dataset splitting to training/cv/testing subsets; done
  • only one equity/currency pair can be traded done
  • no 'skip-frames' implementation within environment; done
  • no plotting features, except if using pycharm integration observer. Not sure if it is suited for intraday strategies. [partially] done
  • making new environment kills all processes using specified network port. Watch out your jupyter kernels. fixed

TODO's and Road Map:

  • [x] refine logic for parameters applying priority (engine vs strategy vs kwargs vs defaults);
  • [X] API reference;
  • [x] examples;
  • [x] frame-skipping feature;
  • [x] dataset tr/cv/t approach;
  • [x] state rendering;
  • [x] proper rendering for entire episode;
  • [x] tensorboard integration;
  • [x] multiply agents asynchronous operation feature (e.g for A3C):
  • [x] dedicated data server;
  • [x] multi-modal observation space shape;
  • [x] A3C implementation for BTgym;
  • [x] UNREAL implementation for BTgym;
  • [x] PPO implementation for BTgym;
  • [ ] RL^2 / MAML / DARLA adaptations - IN PROGRESS;
  • [x] learning from demonstrations; - partially done
  • [ ] risk-sensitive agents implementation;
  • [x] sequential and sliding time-window sampling;
  • [x] multiply instruments trading;
  • [x] docker image; - CPU version, Signalprime contribution,
  • [ ] TF serving model serialisation functionality;

News and updates:

  • 10.01.2019:

  • 9.02.2019:

  • 25.01.2019: updates:

    • lstm_policy class now requires both internal and external observation sub-spaces to be present and allows both be one-level nested sub-spaces itself (was only true for external); all declared sub-spaces got encoded by separate convolution encoders;
    • policy deterministic action option is implemented for discrete action spaces and can be utilised by syncro_runner; by default it is enabled for test episodes;
    • data_feed classes now accept pd.dataframes as historic data dource via dataframe kwarg (was: .csv files only);
  • 18.01.2019: updates:

    • data model classes are under active development to power model-based framework:
      • common statistics incremental estimator classes has been added (mean, variance, covariance, linear regression etc.);
      • incremental Singular Spectrum Analysis class implemented;
      • for a pair of asset prices, two-factor state-space model is proposed
    • new data_feed iterator classes has been added to provide training framework with synthetic data generated by model mentioned above;
    • strategy_gen_6 data handling and pre-processing has been redesigned:
      • market data SSA decomposition;
      • data model state as additional input to policy
      • variance-based normalisation for broker statistics
  • 11.12.2018: updates and fixes:

  • 17.11.2018: updates and fixes:

    • minor fixes to base data provider class episode sampling
    • update to btgym.datafeed.synthetic subpackage: new stochastic processes generators added etc.
    • new btgym.research.startegy_gen_5 subpackage: efficient parameter-free signal preprocessing implemented, other minor improvements
  • 30.10.2018: updates and fixes:

    • fixed numpy random state issue causing replicating of seeds among workers on POSIX os
    • new synthetic datafeed generators - added simple Ornshtein-Uhlenbeck process data generating classes; see btgym/datafeed/synthetic/ and btgym/research/ou_params_space_eval for details;
  • 14.10.2018: update:

    • base reward function redesign -> noticeable algorithms performance gain;
  • 20.07.2018: major update to package:

  • 17.02.18: First results on applying guided policy search ideas (GPS) to btgym setup can be seen here.

    • tensorboard summaries are updated with additional renderings: actions distribution, value function and LSTM_state; presented in the same notebook.
  • 6.02.18: Common update to all a3c agents architectures:

    • all dense layers are now Noisy-Net ones, see: Noisy Networks for Exploration paper by Fortunato at al.;

    • note that entropy regularization is still here, kept in ~0.01 to ensure proper exploration;

    • policy output distribution is 'centered' using layer normalisation technique;

      • all of the above results in about 2x training speedup in terms of train iterations;
  • 20.01.18: Project Wiki pages added;

  • 12.01.18: Minor fixes to logging, enabled BTgymDataset train/test data split. AAC framework train/test cycle enabled via episode_train_test_cycle kwarg.

  • 7.01.18: Update:

    • Major data pipe redesign. Domain -> Trial -> Episode sampling routine implemented. For motivation and formal definitions refer to Section 1.Data of this DRAFT, API Documentation and Intro example. Changes should be backward compatible. In brief, it is necessry framework for upcoming meta-learning algorithms.
    • logging changes: now relying in python logbook module. Should eliminate errors under Windows.
    • Stacked_LSTM_Policy agent implemented. Based on NAV_A3C from DeepMind paper with some minor mods. Basic usage Example is here. Still in research code area and need further tuning; yet faster than simple LSTM agent, able to converge on 6-month 1m dataset.
  • 5.12.17: Inner btgym comm. fixes >> speedup ~5%.

  • 02.12.17: Basic sliding time-window train/test framework implemented via BTgymSequentialTrial() class. UPD: replaced by BTgymSequentialDataDomain class.

  • 29.11.17: Basic meta-learning RL^2 functionality implemented.

  • 24.11.17: A3C/UNREAL finally adapted to work with BTGym environments.

    • Examples with synthetic simple data(sine wawe) and historic financial data added, see examples directory;
    • Results on potential-based functions reward shaping in /research/DevStartegy_4_6;
    • Work on Sequential/random Trials Data iterators (kind of sliding time-window) in progress, start approaching the toughest part: non-stationarity battle is ahead.
  • 14.11.17: BaseAAC framework refraction; added per worker batch-training option and LSTM time_flatten option; Atari examples updated; see Documentation for details.

  • 30.10.17: Major update, some backward incompatibility:

    • BTGym now can be thougt as two-part package: one is environment itself and the other one is RL algoritms tuned for solving algo-trading tasks. Some basic work on shaping of later is done. Three advantage actor-critic style algorithms are implemented: A3C itself, it's UNREAL extension and PPO. Core logic of these seems to be implemented correctly but further extensive BTGym-tuning is ahead. For now one can check atari tests.
    • Finally, basic documentation and API reference is now available.
  • 27.09.17: A3C test_4.2 added:

    • some progress on estimator architecture search, state and reward shaping;
  • 22.09.17: A3C test_4 added:

    • passing train convergence test on small (1 month) dataset of EURUSD 1-minute bar data;
  • 20.09.17: A3C optimised sine-wave test added here.

    • This notebook presents some basic ideas on state presentation, reward shaping, model architecture and hyperparameters choice. With those tweaks sine-wave sanity test is converging faster and with greater stability.
  • 31.08.17: Basic implementation of A3C algorithm is done and moved inside BTgym package.

    • algorithm logic consistency tests are passed;
    • still work in early stage, experiments with obs. state features and policy estimator architecture ahead;
    • check out examples/a3c directory.
  • 23.08.17: filename arg in environment/dataset specification now can be list of csv files.

    • handy for bigger dataset creation;
    • data from all files are concatenated and sampled uniformly;
    • no record duplication and format consistency checks preformed.
  • 21.08.17: UPDATE: BTgym is now using multi-modal observation space.

    • space used is simple extension of gym: DictSpace(gym.Space) - dictionary (not nested yet) of core gym spaces.
    • defined in btgym/
    • raw_state is default Box space of OHLC prices. Subclass BTgymStrategy and override get_state() method to compute alll parts of env. observation.
    • rendering can now be performed for avery entry in observation dictionary as long as it is Box ranked <=3 and same key is passed in reneder_modes kwarg of environment. 'Agent' mode renamed to 'state'. See updated examples.
  • 07.08.17: BTgym is now optimized for asynchronous operation with multiply environment instances.

    • dedicated data_server is used for dataset management;
    • improved overall internal network connection stability and error handling;
    • see example async_btgym_workers.ipynb in examples directory.
  • 15.07.17: UPDATE, BACKWARD INCOMPATIBILITY: now state observation can be tensor of any rank.

    • Consequently, dim. ordering convention has changed to ensure compatibility with existing tf models: time embedding is first dimension from now on, e.g. state with shape (30, 20, 4) is 30x steps time embedded with 20 features and 4 'channels'. For the sake of 2d visualisation only one 'cannel' can be rendered, can be chosen by setting env. kwarg render_agent_channel=0;
    • examples are updated;
    • better now than later.
  • 11.07.17: Rendering battle continues: improved stability while low in memory, added environment kwarg render_enabled=True; when set to False - all renderings are disabled. Can help with performance.

  • 5.07.17: Tensorboard monitoring wrapper added; pyplot memory leak fixed.

  • 30.06.17: EXAMPLES updated with 'Setting up: full throttle' how-to.

  • 29.06.17: UPGRADE: be sure to run pip install --upgrade -e .

    • major rendering rebuild: updated with modes: human, agent, episode; render process now performed by server and returned to environment as rgb numpy array. Pictures can be shown either via matplolib or as pillow.Image(preferred).
    • 'Rendering HowTo' added, 'Basic Settings' example updated.
    • internal changes: env. state divided on raw_state - price data, and state - featurized representation. get_raw_state() method added to strategy.
    • new packages requirements: matplotlib and pillow.
  • 25.06.17: Basic rendering implemented.

  • 23.06.17: alpha 0.0.4: added skip-frame feature, redefined parameters inheritance logic, refined overall stability;

  • 17.06.17: first working alpha v0.0.2.

profile for Andrew Muzikin on Stack Exchange, a network of free, community-driven Q&A sites

Related Awesome Lists
Top Programming Languages
Top Projects

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
Python (807,639
Actions (55,576
Tensorflow (22,344
Trading (5,784
Reinforcement Learning (4,439
Time Series (4,104
Gym (2,506
Unreal Engine (2,252
Deep Reinforcement Learning (1,020
Openai Gym (700
Backtesting (535
A3c (318
Policy Gradient (296
Gym Environment (169
Backtesting Trading Strategies (70
Backtrader (27
Advantage Actor Critic (23
Statistical Arbitrage (9
Algorithmic Trading Library (6
Algoritmic Trading (4
Quantitive Finance (3