deeptendies / legacy-deeptendies-library

Open source stock price forecast bringing quantitative trading to the masses!

---

deeptendies

Bringing quantitative trading to the masses!

The most difficult part of quantitative analysis is getting started. This project has you covered ;)

It is a one-stop shop for obtaining historical data, engineering features, and fitting the data through a pipeline.

Features

Let's open a blank colab and try out this library: https://colab.research.google.com/#create=true

# 1 Liner easy install
!pip install git+https://github.com/stancsz/deeptendies && pip install -r https://raw.githubusercontent.com/stancsz/deeptendies/main/requirements.txt
import deeptendies as dt

# Look and feel of pandas usage & get a pd.DataFrame
df = dt.DataFrame.from_yf('GME')
print(type(df))

# Builtin Pipeline class for mass features processing
pipeline = dt.Pipeline(
    [
        dt.Feature.get_x_high,
        dt.Feature.get_x_low,
        dt.Feature.get_x_ma,
        dt.Feature.get_diff
    ]
)
df = pipeline.run(df=df, x=50, interval='day')
df = pipeline.run(df=df, x=100, interval='day')
df = pipeline.run(df=df, x=200, interval='day')
df = pipeline.run(df=df, x=13, interval='week')
df = pipeline.run(df=df, x=26, interval='week')
df = pipeline.run(df=df, x=52, interval='week')
df[['50_day_ma','200_day_ma']].plot()

happy quanting :)

Additional Features: Feature engineering

Easy to use feature engineering methods

df = dt.Feature.get_x_low(df, x=52, interval='week')
df = get_x_ma(df, x=50, interval='day')

learn more @ deeptendies/feature.py

Jupyter

or use it in a notebook

Development guide

git clone https://github.com/deeptendies/deeptendies.git
pip install -e deeptendies

Credits

• This package is redesigned from the legacy deeptendies package, credits to original authors.
  • @mklasby @bgulseren @KBehairy @hasnil @Karenzhang7717
• This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.

License

• Free software: MIT license
• Documentation: https://deeptendies.readthedocs.io.

legacy-deeptendies-library's Issues

Evalute MLOps Workflow Tools

• Argo - Container-native workflow engine
• MLRun - Orechestration Framework
• KubeFlow - Machine Learning Toolkits for Kubernetes
• Prefect - Workflow managment system
• ZenML - MLOps framework to develop reproducible pipelines

TODO: Provide pros/cons list for each and a recommendaiton for the best solution for our usecase

Evaluate ZenML - MLOps Framework to develop reproducible pipelines

Provide Pros, Cons, and final recommendation(s)

https://zenml.io/

Setup feature store (processed data)

Add get predicted / actual helper to utils class -> shoudl be robust enough to take an arbritary batch size

Start a scraping/data mining project for Marko, pull financial reports from SEC

data scraping / data mining project started at:
https://github.com/deeptendies/mining-sec-edgar

Setup JIRA board and link deeptendies repo

Refs:
https://github.com/integrations/jira
https://support.atlassian.com/jira-cloud-administration/docs/integrate-with-github/

Evaluate BentoML, TensorFlow Serving, and TorchServe for Model serving

Provide Pros, Cons, and final recommendation(s)

https://docs.bentoml.org/en/latest/

Evaluate tslearn

Provide Pros, Cons, and final recommendation(s)

https://github.com/tslearn-team/tslearn

Encapsulate ARIMA for model evaluation/comparison

Setup long term data storage (data lake)

add gridsearch with autokeras

def of done

• make a makeshift poc
• make it more robust

grid search on

1. dataset range
2. epoch
3. batch size

each autokeras timeseries trial should log the performance

Evalute MLRun - Orchestration Framework

Provide Pros, Cons, and final recommendation(s)

Evalute Prefect - Workflow management system

Provide Pros, Cons, and final recommendation(s)

https://docs.prefect.io/

Evaluate low-latency data serving options (Redis, GCP MemStore, etc.)

Consider pros, cons, and cost. Make recommendaiton

Spin up a celery scheduled tasks server

Spin up a scheduled task server

• pulls data daily
• save data to csv / jsonline buckitized
• build docker image for it

similar to update data in a public google sheet like this

boilerplate celery app: https://github.com/stancsz/celery-scheduler-app

Evalute KubeFlow - Machine Learning Toolkits for Kubernetes

Provide Pros, Cons, and final recommendation(s)

https://www.kubeflow.org/

Evaluate AutoArima , pmdarima

wrote a minimalist/lite fork of deeptendies. just single column autoarima (R's autoarima's python rewrite)
https://pypi.org/project/pmdarima/

        stock='gme',
        start='2021-05-10',
        end='2021-05-27',
        forecast_days=5,
        filter=['High']

>> 
/usr/bin/python3.8 /home/stan/github/deeptendies-lite/main.py
[272.1088269  279.6325006  287.15617429 294.67984799 302.20352168]

Process finished with exit code 0

Evalute Argo - Container-native workflow engine

Provide Pros, Cons, and final recommendation(s)

https://argoproj.github.io/

Explore darts https://unit8co.github.io/darts/index.html

Implement naive backtest algorithm for existing LSTM models

EG., make a weekly buy/sell pair based on next 5-day high/low price. Do not work about modelling market conditions yet, we just want to see a theoretical maximum ROI

Evaluate GlueonTS

Evaluate https://ts.gluon.ai/
figure out if it is suitable for statistical arbitrage use case, write some blurb about it for sharing the knowledge

• read the documentation
• follow the quick start
• run the quick start and share code
• write a blurb about the evaluation, write on medium
• pros/cons list, what does it looks like it's best for

Implement technical indicators

rescope needed
We came across this library https://twopirllc.github.io/pandas-ta/, it seems like it has most of the technical indicators we need. let's add the library to our stack for now and only write our own ta logics if needed.

Let's think about how can we "grid search" a couple of indicators for each dataset (stocks). I think we can try something like:

• correlation analysis
• strategy evaluation

We can pick a couple of clusters, then based on each cluster, we pick the most correlated metric and use them as feature columns to train our timeseries.