/LibtorchTrade

RL BTC trader written in C++ using libtorch

Primary LanguageC++MIT LicenseMIT

LibtorchTrade

Reinforcement learning trader for BTC written in C++ using libtorch

Guiding principles

  • Fast
    • Written purely in C++.
    • Support multi-threading training.
    • Since the daily historical data is a small dataset(4MB), it can easily fit into the cache. So it is at least twice faster to train using CPU than GPUs.
    • A sample training log, it is just 20 seconds training.
    2020-11-18 22:13:36 +0000 CSVTensor INFO load data from ../data/btc.csv
2020-11-18 22:13:37 +0000 Trainer INFO feature_add_rnn start
2020-11-18 22:13:38 +0000 Trainer DEBUG 0 train_reward=12.9045 test_reward=1.62863
...
2020-11-18 22:13:51 +0000 Trainer DEBUG 20 train_reward=204.805 test_reward=2.5557
2020-11-18 22:13:51 +0000 Trainer INFO  save model: epoch=20 path=../models_train/btc/1d/PPOSampler_AttGru_1d_1rnn32_linear32_drop0.2_nlen30_Adam_8-9-104-114-122-129-0_prod2.5557_epo20_1605737631_roulette.pt
  • Effective
    • Here is a test result from 1 model during Dec 2019 and Dec 2020. I trained it with data from March 2013 to Nov 2019: alt text

Build

  • You need to install libtorch C++ first, then run the commands below:
mkdir build
cd build
cmake -DCMAKE_PREFIX_PATH=<path to libtorch> ..
make

Training

  • A simple heuristic feature selection algorithm is implemented. To train with this algorithm:
train feature_add