Skip to content
A messy representation of some of my deep learning experiments and learning 2016/2017
Jupyter Notebook Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.DS_Store
.gitignore
README.md
conv1d_failed_autoencoder.py
distal-AE-july6-240pm-LSTM-99-0.83.h5
load_data.py
model_performance_evaluation.py
proximal-AE-july6-131pm-LSTM-83-0.84.h5
rectangle_here-AE-july6-301pm-Conv1D-39-0.63.h5
the-internet-of-forex-05.15.17-1217pm-Copy1.ipynb
the-internet-of-forex-06.25.17.404pm-LSTM-rabbit-hole-round2.ipynb
the-internet-of-forex-06.29.17-LSTM-409pm.ipynb
the-internet-of-forex-07.03.17.910pm-Copy1-LSTM-4096.ipynb
the-internet-of-forex-07.04.17-softmax-temperature-experiments-Copy1.ipynb
the-internet-of-forex-07.05.17-full-pipeline-524pm.ipynb
the-internet-of-forex-07.05.17-rectangle-here-training-Copy1.ipynb
the-internet-of-forex-07.05.17-retrain-all-models-Copy1.ipynb
the-internet-of-forex-07.05.17.1009am-LSTM-for-proximal-and-distal.ipynb
the-internet-of-forex-07.06.17-distal-339pm.ipynb
the-internet-of-forex-07.06.17-proximal-339pm.ipynb
the-internet-of-forex-07.07.17-proximal-distal-TOGETHER.ipynb
the-internet-of-forex-07.08.17-rectangle-here-Copy1-forest-feature-selection.ipynb
the-internet-of-forex-07.08.17-rectangle-here.ipynb
tw_conv1d_best_jun30.py
weightnorm.py
zone feature engineering 07 12 17.png

README.md

Experiments in machine learning and deep learning with Forex

This repo is intended as some sort of a (very disorganized, highly unacademic, and somewhat embarrassing) public archive of a year long experimentation of applying machine learning techniques to FX charts.

The idea was as follows: Can we use breakthroughs in deep learning to replicate the behavior of an experienced day trader?

I started by collecting some data from my friend Rudy's day-trading platform, including getting some export scripts in place.

Then, I tried to apply some simple feed forward networks.

Then, I tried random forests, and about everything else relevant sklearn.

Then, I though "Why not try applying some breakthrough papers in image recognition to the problem of 'identifying' these shapes, but represented numerically instead ?"

Then, came a whole lot of work, specifically with convolutional nets, LSTMs, and autoencoders. Everytime I found a new paper that seemed relevant I would try to implement it for my problem.

In the end, a lot of feature engineering (with help from my awesome tutor, Joe) and a pretty simple Conv1D architecture based loosely on a paper related to mammogram identification achieved the highest F1 score and was actually pretty useful.

About 5% of the code I wrote is contained herein.

If you find it useful, let me know!

You can’t perform that action at this time.