Deep Quant

by Euclidean Technologies, LLC

On a periodic basis, publicly traded companies are required to report fundamentals: financial data such as revenue, operating income, debt, among others. These data points provide some insight into the financial health of a company.

This repository contains a set of deep learning tools for forecasting company future fundamentals from historical fundamentals and other auxiliary data such as historical prices and macro economic data.

Installation and Setup

Clone repo, setup environment, and install requirements:

$ git clone https://github.com/euclidjda/deep-quant.git
$ cd deep-quant
$ export DEEPQUANTROOT=pwd
$ pip install -r requirements.txt

You may also want to put DEEPQUANTROOT in your shell initialization file such .bash_profile so that it does not need to be defined every time you start a shell. For example, you could run the following from within the deep-quant directory:

$ echo "export DEEPQUANTROOT="pwd >> ~/.bash_profile

Preparing the Data

If you have access to Wharton Research Data Services (WRDS) through your academic institution, go here to learn how to create a dataset for deep-quant.

WRDS instructions for creating a deep-quant data file for learning and forecasting.

**Do not use models built with the dataset described below for actual trading or investing. This is a freely available dataset assembled from freely available sources and contains errors such as look-ahead bias and survivorship bias.**

Data is passed to deep_quant.py as a space-delimited flat file. If you do not have access to a commercial or academic dataset and you would like to test this code, we have provided a "open dataset" for this purpose. Again this dataset should be use for testing purposes only. To obtain this file, run the command:

$ python scripts/build_datfile.py

This will create a datasets/open-dataset.dat file.

Building Models

You can train deep quant on a neural network of a particular type and of a particular architecture with several other hyperparameters on a particular dataset by first defining all of these things on a config file, and then specifying that config file as the point of reference when running deepquant.py. Consider, for example, how deepquant is run on open-dataset.dat, as specified by config/system-test.conf:

$ python scripts/deep_quant.py --config=config/system-test.conf --train=True

This will load the corresponding data and cache it in batches in a directory called _bcache, and will save model checkpoints in a directory called chkpts-system-test (both of these directories will be created automatically).

A couple of notes about config files:

* The user can specify a .dat file to use through the --datafile and the

data_dir options (note that the latter is datasets by default).

* financial_fields is a range of columns, and should be specified as a

string joining the first and last columns of the .dat file that the user

wants to forecast (for example: saleqttm-ltqmrq).

* aux_fields is similarly also a range of columns that is equivalently

specified. Note, however, that these fields are strictly features; they are

not part of what the model is trained to predict.

Generating Forecasts

To generate forecasts for the companies in the validation set, deep_quant.py must be run with the --train option set to False. For example:

$ python scripts/deepquant.py --config=config/system-test.conf --train=False --prettyprintpreds=True --mseoutfile=mse-data.txt > forecasts.txt

That'll produce a file called forecasts.txt with the predicted values for every financial feature at every timestep.

Hyper-parameter Search

The deep-quant repository contains tools for performing hyper-parameter searches. A hyper-parameter search requires a datafile and a configuration template. A configuration template takes the format of the learning configuration file with the exception that each configation parameter has multiple values, seperated by spaces (for example an, see hyper-search.conf). The hyper-parameter search algorithms uses the template to definine the hyper-parameter search space (i.e., all possible parameter combinations). A user may specify one of two search algorithms: grid_search or genetic. To experiment with hyper-parameter search, execute the following from the deep-quant directory:

$ mkdir search
$ cp config/hyper-search.conf search/.
$ cd search
$ python $DEEPQUANT/scripts/hyperparamsearch.py --template=hyper-search.conf --searchalgorithm='genetic'

Uncertainty Quantification (UQ)

Forecast uncertainty is obtained by using UQ compatible deep learning models available in the models directory. Examples of the UQ parameters to be used in the config file (eg system-test.conf) are as follows:

--UQ                    True --nn_type               DeepLogLikelihoodUQModel --UQmodeltype         MVE --dfdirname            outputsdfs --keepprobpred        0.7

will contain the corresponding output dataframes for prediction, data noise variance and errors. Total variance is the sum of data noise variance (output of the NN) and model variance. Model variance is calculated by performing the same experiment multiple times with different random seed. Confidence Intervals can be calculated using the predictions and the total variance.

Contributors and Acknowledgement

This repository was developed and is maintained by Euclidean Technologies, LLC. Individual core contributors include John Alberg, Zachary Lipton, Lakshay Kumar Chauhan, and Ignacio Aranguren.

License

This is experimental software. It is provided under the [MIT license][mit], so you can do with it whatever you wish except hold the authors responsible if it does something you don't like.

[mit]: http://www.opensource.org/licenses/mit-license.php