lauraperge / blackbox-trading-cnn

Time Series Classification with Convolutional Neural Network: Automated Trading by Pattern Recognition (Master's Thesis)

blackbox-trading-CNN

Time Series Classification with Convolutional Neural Network: Automated Trading by Pattern Recognition (Master's Thesis)

Prequisites

All parts are written in Python 3 (except report which is in LaTex) the requirements can be installed using:

pip install -r requirements.txt

For the CNN and Data preparation I used Jupyter notebooks.

Content

data

All downloaded datasets, trial code for data retrivals via API. /data_raw/* has all raw datasets used to create the inputs

/cnn_input/* has all folders that will hold all inputs for the CNN given the data-check-and-prep-alldata-2Y.ipynb notebook is ran.

/competing-strategies-signals/* contains the signals for the test assets based on the RSI and BB algorithmic trading trading_strategies

/logs/* gets all logs from data-check-and-prep-alldata-2Y.ipynb

notebooks

Jupyter notebooks of CNN versions and Data Preparation

• data-check-and-prep-alldata-2Y uses prewritten modules (See utils) to preliminarily check/modify the data, label and transform it to be ready for the CNN. It visualizes the input/output data and labelling. Its input is any raw dataset (data/data_raw), its output goes into data/cnn_input/ and serves as input for the CNN notebooks.

• Google Colab codes/* holds all asset-specific and universal model codes which are trained and tested in 3 different training periods (dates for period 2 to be checked in notebooks that are not ran on period two in case of general use.) CNN, Keras

• financial-evaluation takes all results from the Results folder and competing-strategies-signals folder and returns financial performance measurements and figures

• competing-strategies creates trading signals for the same datasets given RSI and BB algorithmic trading strategy

• classification&time-evaluation takes all results from the Results folder and returns classification performance metrics

• /Output/ contains logs from the CNN, later will contain other evaluation outputs as well

results tables

Contains all result outputs from financial-evaluation.ipynb* and classification&time-evaluation.ipynb, contains general results table

report

LaTeX report

utils

All custom modules for obtaining, examining, preparing, cleaning, labelling, and transforming data into images, as well as a code to prepare input for the tensorflow CNN

• data cleaning: read in raw data, turn into pandas time series, resampling if needed, missing value checks, reporting feature

• get_data: functions to retrieve historical cryptocurrency data, stock prices and fx prices via API connections (sources: bitfinex, yahoo, fred)

• labelled_image_preparation: takes vector as input, returns the transformed version as images of given size according to given transformation strategy/strategies with relevant trading labels that serves as an input for the tensorflow CNN

labels

labelling strategy for the time series

• trading_strategies: local min-max implementation

transform

• gramian angular field: for creating one ore more gramain angular summation/difference fields from a vector representing a series of observations

• markov_transition_field: for creating one or more markov transition fields from a vector representing a series of observations (as per https://www.researchgate.net/publication/275970614_Encoding_Time_Series_as_Images_for_Visual_Inspection_and_Classification_Using_Tiled_Convolutional_Neural_Networks)

• recurrence_plot: for creating one or more recurrence plots from a vector representing a series of observations (as per https://arxiv.org/abs/1710.00886)

visualize

• ts_with_markers: show trading strategy Buy/Sell points on time series