duckbill / time2image-ERP-PREDICTION-CONTEST

Team Jonas contribution to ERP Prediction Contest, February 15, 2019 - May 15, 2019

ERP PREDICTION CONTEST

Winning contribution
The UCSB Department of Statistics & Applied Probability and the Center for Financial Mathematics and Actuarial Research are partnering with Hull Tactical to run a data science contest on predicting the S&P 500 returns.

Link to competition: https://ucsb-erp-contest.herokuapp.com/
Slides can be found at: https://docs.google.com/presentation/d/1zC-y9jtTpfaNGIR7_5C1oHYKibpa8ITIv0sUOjxE1qk/edit?usp=sharing

Victory Image :)

Tools and Methods Used In Project

• Pytorch
• Fastai
• Multiple and Single Input Convolutional Neural Networks
• Image Creation using Gramian Angular Fields

Aim

Images are created using Gramian Angular Field [1] of time series data. Using Python, Fastai and Pytorch a Convolutional Neural Network is created to take multiple image imputs.

[1] Image representation created using Gramian Angular Field described by Zhiguang Wang and Tim Oates (Jun 2015) here: https://arxiv.org/pdf/1506.00327.pdf

Description of files:

• Jonas ERP Contest.pdf: Complete report on project.
• 4 percentage+BB+MA20050diff+MACD.html: The same notebook as 4 percentage+BB+MA20050diff+MACD.ipynb where all steps has bee commented.
• dataset.csv: The data for the competition, can be found at https://ucsb-erp-contest.herokuapp.com/
• imagecreater.py: Adds relevant features to the data set and creates image representation of time series data
• ResNet50 CLOSE.ipynb: Creates, trains and tests a CNN for single input: CLOSE.
• ResNet50 percentage.ipynb: Creates, trains and tests a CNN for single input: percentage.
• 2 CLOSE+MA20050diff.ipynb: Creates, trains and tests a CNN for 2 inputs: CLOSE and MA20050diff.
• 2 percentage+BB.ipynb: Creates, trains and tests a CNN for 2 inputs: percentage and BB.
• 2 percentage+RSI.ipynb: Creates, trains and tests a CNN for 2 inputs: percentage and RSI.
• 4 CLOSE+percentage+RSI+BBdiff.ipynb: Creates, trains and tests a CNN for 4 inputs: CLOSE, percentage, RSI and BB.
• 4 percentage+BB+MA20050diff+MACD.ipynb: Creates, trains and tests a CNN for 4 inputs: percentage, BB, MA20050diff, MACD. All steps has bee commented.

Note

The file containing the images is too large to upload. The images can be reconstructed using the imagecreater.py and dataset.csv.