replication of the paper "Deep Learning Statistical Arbitrage",available at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3862004 and https://arxiv.org/abs/2106.04028.

• Fama-French factor models for residuals computation
• Various machine learning models ( OUFFN, CNNTransformer, and FourierFFN) to construct and optimize trading strategies
• With the goal of maximizing risk-adjusted returns(Sharpe ratio and mean-var)
• Data from S&P 500, by yfinance to collect 10 stocks.

The project is organized as follows:

• famafrench.ipynb contains the code for data collecting and creating residuals from raw data

• train_test.py contains the code for training a trading policy model and simulating trading

• train_test_2.py contains the modified code from train_test.py (add excess returns to optimize the objective)

• run_train_test.ipynb is a user interface to train_test.py which deals with configuration, saving results, etc

• preprocess.py contains functions for preprocessing residual time series data into a form usable by a trading policy model

• utils.py contains helpful functions used throughout

• results should contains folders, which will contain the results of and plots for trading policy model tests conducted by run_train_test.py

• models contains code for trading policy models

• config contains configuration files which define various tests of trading policy models on residual time series

• residuals contains data generated from famafrench.ipynb

• First to run famafrench.ipynb to generate residuals, and will be saved in the residuals folder

• Then run run_train_test.ipynb

• If you have access to a GPU, ensure PyTorch detects it properly, and set the device to cuda in the configuration files.

• You can also adjust the batch_size and num_epochs in the configuration files to speed up training.

• Make sure you have folders such as /results/Unknown, /results/cumulative_returns, /results/turnovers ,/results/short_proportions, /results/Checkpoint at first
• Make sure __init__.py exists in the models folder，to ensure Python treats the models folder as a package

• Function 'StdBackward0' returned nan values in its 0th output.

  Make every std=torch.std (__ + 1e-8 ) + 1e-8