initial commit

.gitignore

@@ -0,0 +1,145 @@
+# Project-specific files
+daily-returns.csv
+*esiduals*.npy
+*.npz
+credentials.yaml
+.vscode/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+# Usually these files are written by a python script from a template
+# before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+# For a library or package, you might want to ignore these files since the code is
+# intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+# However, in case of collaboration, if having platform-specific dependencies or dependencies
+# having no cross-platform support, pipenv may install dependencies that don't work, or not
+# install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/

README.md

@@ -0,0 +1,42 @@
+Deep Learning Statistical Arbitrage
+===================================
+
+This repo contains the official code for our paper *Deep Learning Statistical Arbitrage*, available at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3862004 and https://arxiv.org/abs/2106.04028.
+
+## Quickstart
+
+To test a trading policy model on a residual time series, use `run_train_test.py`. 
+This file exports a function, `run()`, which can be imported and used in e.g. a
+grid search, or run from the command line. Command line usage will suit most users.
+To run from the command line, use
+```
+python3 run_train_test.py -c configs/config_name_here.yaml
+```
+where `config_name_here.yaml` is a configuration file from the `configs` folder.
+You can write your own configuration file to edit hyperparameters and other 
+settings for the trading test. See `run_train_test.py` for other command line options.
+
+## Structure
+
+This repo is organized as follows:
+- `train_test.py` contains the code for training a trading policy model and simulating trading.
+- `run_train_test.py` is a user interface to `train_test.py` which deals with configuration, logging, saving results, etc.
+- `preprocess.py` contains functions for preprocessing residual time series data into a form usable by a trading policy model
+- `data.py` contains miscellaneous functions for altering residual time series data
+- `config` contains configuration files which define various tests of trading policy models on residual time series
+- `data` should contain raw input data used to create residuals
+- `factor_models` contains code for creating residuals from raw input data
+- `residuals` stores residual time series data sets created by the code in `factor_models`
+- `models` contains code for trading policy models
+- `results` will contain the results of and plots for trading policy model tests conducted by `run_train_test.py`
+- `logs` will contain logs for runs of models and factor models
+- `tools` should contain miscellaneous code for interpreting and exploring results and saved models
+- `utils.py` contains helpful functions used throughout
+
+## Generating residuals
+
+To create residuals, first ensure that input data is present in the `data` directory, then run `run_factor_model.py`, providing the name of a factor model in the `factor_models` directory, e.g.
+```
+python3 run_factor_model.py -m factor_model_name_here
+```
+Generated residuals for the factor model will be saved in the `residuals` folder.

configs/cnntransformer-full.yaml

@@ -0,0 +1,51 @@
+# Major parameters
+mode: "test" # can be 'test' or 'estimate'
+results_tag: "" # optional; try not to use underscores in this tag, use dashes instead
+debug: False # set to True to turn on debug logging and file naming
+# Model parameters
+model_name: "CNNTransformer" # name of a class defined in models folder and initialized in model folder's __init__.py
+model: { # contains parameter settings for __init__() function of class with name `model_name`
+ lookback: 30, # number of days of preprocessed residual time series to feed into model
+ dropout: 0.25,
+ filter_numbers: [1,8],
+ filter_size: 2, 
+ attention_heads: 4,
+ hidden_units_factor: 2, # multiplicand of last item in `filter_numbers`; determines number of hidden units (e.g. 2*8 = 16)
+ # hidden_units: 16, # use either hidden_units or hidden_units_factor, but not both
+ normalization_conv: True, # normalize convolutions or not
+ use_transformer: True,
+ use_convolution: True,
+}
+# Data parameters
+preprocess_func: "preprocess_cumsum" # name of a function defined in preprocess.py
+use_residual_weights: True # use residual composition matrix to compute turnover, short proportion, etc.
+cap_proportion: 0.01 # defines asset universe: 0.01 corresponds to a residual data set
+factor_models: { # number of factors per residual time series to test, for each factor model
+ "IPCA": [5],
+ "PCA": [5],
+ "FamaFrench": [5],
+}
+perturbation: { # perturbation of residual time series by noise is optional, leave empty or comment out entirely to disable
+ # "noise_type" : "gaussian",
+ # "noise_mean" : 0.0,
+ # "noise_std_pct" : 2,
+ # "noise_only" : False,
+ # "per_residual" : True,
+}
+# Training parameters
+num_epochs: 100
+optimizer_name: "Adam" # see PyTorch docs for potential optimizers
+optimizer_opts: { # see PyTorch docs for optimizer options
+ lr: 0.001
+}
+batch_size: 125
+retrain_freq: 125 # if mode=='estimate', this is the number of obs used to form a test set (chronologically after the training set)
+rolling_retrain: True # set to False for no rolling retraining (i.e. train once, test for all data past training set)
+force_retrain: True # force the model to be trained, even if existing weights for the model are saved on disk
+length_training: 1000 # size of rolling training window in trading days
+early_stopping: False # employ early stopping or not
+objective: "sharpe" # objective function: 'sharpe' or 'meanvar' or 'sqrtMeanSharpe'
+# Market frictions parameters
+market_frictions: False # enable or disable
+trans_cost: 0 # cost in bps per txn side per equity, e.g. 0.0005
+hold_cost: 0 # cost in bps for short positions per equity per day, e.g. 0.0001

configs/fourierffn-full.yaml

@@ -0,0 +1,44 @@
+# Major parameters
+mode: "test" # can be 'test' or 'estimate'
+results_tag: "" # optional; try not to use underscores in this tag, use dashes instead
+debug: False # set to True to turn on debug logging and file naming
+# Model parameters
+model_name: "FourierFFN" # name of a class defined in models folder and initialized in model folder's __init__.py
+model: { # contains parameter settings for __init__() function of class with name `model_name`
+ lookback: 30, # number of days of preprocessed residual time series to feed into model
+ dropout: 0.25,
+ hidden_units: [30,16,8,4], # must start with 30
+}
+# Data parameters
+preprocess_func: "preprocess_fourier" # name of a function defined in preprocess.py
+use_residual_weights: False # use residual composition matrix to compute turnover, short proportion, etc.
+cap_proportion: 0.01 # defines asset universe: 0.01 corresponds to a residual data set
+factor_models: { # number of factors per residual time series to test, for each factor model
+ "IPCA": [5],
+ "PCA": [5],
+ "FamaFrench": [5],
+}
+perturbation: { # perturbation of residual time series by noise is optional, leave empty or comment out entirely to disable
+ # "noise_type" : "gaussian",
+ # "noise_mean" : 0.0,
+ # "noise_std_pct" : 2,
+ # "noise_only" : False,
+ # "per_residual" : True,
+}
+# Training parameters
+num_epochs: 100
+optimizer_name: "Adam" # see PyTorch docs for potential optimizers
+optimizer_opts: { # see PyTorch docs for optimizer options
+ lr: 0.001
+}
+batch_size: 125
+retrain_freq: 125 # if mode=='estimate', this is the number of obs used to form a test set (chronologically after the training set)
+rolling_retrain: True # set to False for no rolling retraining (i.e. train once, test for all data past training set)
+force_retrain: True # force the model to be trained, even if existing weights for the model are saved on disk
+length_training: 1000 # size of rolling training window in trading days
+early_stopping: False # employ early stopping or not
+objective: "sharpe" # objective function: 'sharpe' or 'meanvar' or 'sqrtMeanSharpe'
+# Market frictions parameters
+market_frictions: False # enable or disable
+trans_cost: 0 # cost in bps per txn side per equity, e.g. 0.0005
+hold_cost: 0 # cost in bps for short positions per equity per day, e.g. 0.0001

configs/ouffn-full.yaml

@@ -0,0 +1,44 @@
+# Major parameters
+mode: "test" # can be 'test' or 'estimate'
+results_tag: "" # optional; try not to use underscores in this tag, use dashes instead
+debug: False # set to True to turn on debug logging and file naming
+# Model parameters
+model_name: "OUFFN" # name of a class defined in models folder and initialized in model folder's __init__.py
+model: { # contains parameter settings for __init__() function of class with name `model_name`
+ lookback: 30, # number of days of preprocessed residual time series to feed into model
+ dropout: 0.25,
+ hidden_units: [4,4,4,4], # must start with 4, the OU signal length
+}
+# Data parameters
+preprocess_func: "preprocess_ou" # name of a function defined in preprocess.py
+use_residual_weights: False # use residual composition matrix to compute turnover, short proportion, etc.
+cap_proportion: 0.01 # defines asset universe: 0.01 corresponds to a residual data set
+factor_models: { # number of factors per residual time series to test, for each factor model
+ "IPCA": [5],
+ "PCA": [5],
+ "FamaFrench": [5],
+}
+perturbation: { # perturbation of residual time series by noise is optional, leave empty or comment out entirely to disable
+ # "noise_type" : "gaussian",
+ # "noise_mean" : 0.0,
+ # "noise_std_pct" : 2,
+ # "noise_only" : False,
+ # "per_residual" : True,
+}
+# Training parameters
+num_epochs: 100
+optimizer_name: "Adam" # see PyTorch docs for potential optimizers
+optimizer_opts: { # see PyTorch docs for optimizer options
+ lr: 0.001
+}
+batch_size: 125
+retrain_freq: 125 # if mode=='estimate', this is the number of obs used to form a test set (chronologically after the training set)
+rolling_retrain: True # set to False for no rolling retraining (i.e. train once, test for all data past training set)
+force_retrain: True # force the model to be trained, even if existing weights for the model are saved on disk
+length_training: 1000 # size of rolling training window in trading days
+early_stopping: False # employ early stopping or not
+objective: "sharpe" # objective function: 'sharpe' or 'meanvar' or 'sqrtMeanSharpe'
+# Market frictions parameters
+market_frictions: False # enable or disable
+trans_cost: 0 # cost in bps per txn side per equity, e.g. 0.0005
+hold_cost: 0 # cost in bps for short positions per equity per day, e.g. 0.0001

configs/rawffn-full.yaml

@@ -0,0 +1,44 @@
+# Major parameters
+mode: "test" # can be 'test' or 'estimate'
+results_tag: "" # optional; try not to use underscores in this tag, use dashes instead
+debug: False # set to True to turn on debug logging and file naming
+# Model parameters
+model_name: "RawFFN" # name of a class defined in models folder and initialized in model folder's __init__.py
+model: { # contains parameter settings for __init__() function of class with name `model_name`
+ lookback: 30, # number of days of preprocessed residual time series to feed into model
+ dropout: 0.25,
+ hidden_units: [30,16,8,4], # must start with lookback
+}
+# Data parameters
+preprocess_func: "preprocess_cumsum" # name of a function defined in preprocess.py
+use_residual_weights: False # use residual composition matrix to compute turnover, short proportion, etc.
+cap_proportion: 0.01 # defines asset universe: 0.01 corresponds to a residual data set
+factor_models: { # number of factors per residual time series to test, for each factor model
+ "IPCA": [5],
+ "PCA": [5],
+ "FamaFrench": [5],
+}
+perturbation: { # perturbation of residual time series by noise is optional, leave empty or comment out entirely to disable
+ # "noise_type" : "gaussian",
+ # "noise_mean" : 0.0,
+ # "noise_std_pct" : 2,
+ # "noise_only" : False,
+ # "per_residual" : True,
+}
+# Training parameters
+num_epochs: 100
+optimizer_name: "Adam" # see PyTorch docs for potential optimizers
+optimizer_opts: { # see PyTorch docs for optimizer options
+ lr: 0.001
+}
+batch_size: 125
+retrain_freq: 125 # if mode=='estimate', this is the number of obs used to form a test set (chronologically after the training set)
+rolling_retrain: True # set to False for no rolling retraining (i.e. train once, test for all data past training set)
+force_retrain: True # force the model to be trained, even if existing weights for the model are saved on disk
+length_training: 1000 # size of rolling training window in trading days
+early_stopping: False # employ early stopping or not
+objective: "sharpe" # objective function: 'sharpe' or 'meanvar' or 'sqrtMeanSharpe'
+# Market frictions parameters
+market_frictions: False # enable or disable
+trans_cost: 0 # cost in bps per txn side per equity, e.g. 0.0005
+hold_cost: 0 # cost in bps for short positions per equity per day, e.g. 0.0001