This github respository is for our ICAIF 2022 paper "Theoretically Motivated Data Augmentation and Regularization for Portfolio Construction," see the full version of our paper on arxiv: https://arxiv.org/abs/2106.04114
We include a demonstration implementation and application of the proposed method to learning a portfolio during the 2020 Market Crash. See the file demonstration_market_crash.ipynb
The code in main.py can be used to grid search over different parameters for all stocks.
To obtain the S&P500 data we used, change the directory to ./data, use the following command, and extract the csv file:
wget https://github.com/pfnet-research/Finance_data_augmentation_ICAIF2022/releases/download/data/sp500.zip
The following figures shows portfolio constructed by our neural network model of MSFT (Microsoft) during the 2020 stock market crash (see https://en.wikipedia.org/wiki/2020_stock_market_crash). The vertical dashed line shows the data of the market crash.
The price of the stock:
The constructed portfolio:
As we see, the model avoids the crash quite well. The demonstration code for training and testing this model is here in the jupyter notebook file demonstration_market_crash.ipynb.