Continual Learning on financial time series

Code repo for thesis in Continual Learning @ Axyon AI

Model is customizable, for further information refer to

python ./main.py -h

Input data can also be pre-processed with --processing:

• --none
• --difference
• --indicators (Project is tested with this option)

An example of BOCD is available at notebook/chp_analysis.ipynb and running python ./main.py --split

Package and dependencies:

First, install the BOCD package via

pip install -e .

in detection repo

then install dependencies via

pip install -r requirements.txt

in main project repo

PyTorch and TA-Lib are also required, please visit the official pages

Model:

• By default, the DL backbone used is a simple MLP with Dropout
• In addition a CNN with 1-D Convolution with L1 Norm regulariation might be used (to be fixed and tested) with the argument --cnn

Online learning:

Model can be executed for both regression and classification (regression not working at the moment)

Online training:

python ./main.py  

or

python ./main.py  --model online

Other CL methods

With the option --model can be trained (and tested) with the following models:

• Averaged GEM
• Avergaed GEM with Reservoir
• Dark Experience Replay
• Elastic Weight Consolidation
• Experience Replay
• Gradient Episodic Memory
• Synaptic Intelligence

Evaluation

Wiht the argument --evaluate each model can be tested each epoch for both current and previous tasks with a final recap .csv file.

TO DO:

• Finish regularization for CNN
• Classification on % of outscore/outperform and not on price
• Test with different sequence timestep (actually 30 days of observation and prediction 30 days later)