torchquantum

TorchQuantum is a backtesting framework that""" integrates the structure of PyTorch and WorldQuant's Operator for efficient quantitative financial analysis.

Contents

• Installation
• Features
• Contribution

Installation

for Unix:

cd /path/to/your/directory
git clone git@github.com:nymath/torchqtm.git
cd ./torchqtm

Before running examples, you should compile the cython code.

python setup.py build_ext --inplace

Now you can run examples

python ./examples/main.py

If you are not downloading the dataset, then you should

cd ./examples
mkdir largedata
cd ./largedata
wget https://github.com/nymath/torchqtm/releases/download/V0.1/stocks_f64.pkl.zip
unzip stocks_f64.pkl.zip
rm stocks_f64.pkl.zip
cd ../
cd ../
git checkout dev

As for the backtesting dataset, we use the bundle provided by ricequant. We have wrapped the code into Makefile, you can just run the following command to download the bundle.

make rqalpha_download_bundle

for windows: We highly recommend you to use WSL2 to run torchquantum.

Examples

alpha mining

You can easily create an alpha through torchquantum!

import torchqtm.op as op
import torchqtm.op.functional as F


class NeutralizePE(op.Fundamental):
    def __init__(self, env):
        super().__init__(env)
        self.lag = op.Parameter(5, requires_optim=False, feasible_region=None)

    def forward(self):
        self.data = F.divide(1, self.env.PE)
        self.data = F.winsorize(self.data, 'std', 4)
        self.data = F.normalize(self.data)
        self.data = F.group_neutralize(self.data, self.env.Sector)
        self.data = F.regression_neut(self.data, self.env.MktVal)
        self.data = F.ts_mean(self.data, self.lag)
        return self.data

• F is library that contains the operators defined by WorldQuant.
• op.Fundamental implies the NeutralizePE belongs to fundamental alpha.
• self.lag is the parameter of rolling mean, which can be optimized through grid search.

backtesting

Here we create a buy and hold strategy for illustration.

from torchqtm.edbt.algorithm import TradingAlgorithm
from torchqtm.assets import Equity

class BuyAndHold(TradingAlgorithm):
    def initialize(self):
        self.safe_set_attr("s0", Equity("000001.XSHE"))
        self.safe_set_attr("count", 0)

    def before_trading_start(self):
        pass

    def handle_data(self):
        if self.count == 0:
            self.order(self.s0, 10000)
        self.count += 1

    def analyze(self):
        pass

Features

• High-speed backtesting framework (most of the operators are implemented through cython)
• A revised gplearn library that is compatible with Alpha mining.
• CNN and other state of the art models for mining alphas.
• Event Driven backtesting framework is available.

Contribution

For more information, we refer to Documentation.

Join us

If you are interested in quantitative finance and are committed to devoting your life to alpha mining, you can contact me through WeChat at Ny_math.

References

quantopian/alphalens

quantopian/zipline