Airain is a research project aiming at building an AI able to suggest optimal portfolio allocation for day trading (currently for stocks and cryptocurrencies). On top of that, it is able to open and close its recommended portfolio.

• 📡 Data retrieval using Yahoo Finance API (code for Intrinio's API are also proposed)
• 🏋️ Model training (see the architecture below)
• 🤺 Model backtesting
• 🦾 Automated, fast order placing in Trading 212 using Selenium for Firefox
• 🤖 Background-running trader for automated data updating and opening/closing positions
• 📈 Graphical interface using Dash (under development)

Clone the repo :

git clone https://github.com/jdenes/airain.git

Then install requirements:

pip install -r requirements.txt

A Trading 212 account and credentials (a free version allows you to trade using fake money). See trading212_example.cfg files in /resources/ to understand how to fill credentials. Yours should be place in a new trading212.cfg file in the same folder.

All codes are stored in /src/ and should be executed there. The only interaction should be with main.py, where high-level commands are defined. Comment or uncomment functions in "__main__" and execute (better version with argparse to come):

cd /src/
python main.py

This repo contains the result of more than two years of experiments. It consists of a final chosen model architecture, but also a lot of unused material. We present then both here, with their designated files.

Heavily inspired by the paper Deep Learning for Portfolio Optimization, the intuition is to use a neural network which output shape is the number of asset plus one, representing the cash bias. The reward used is the obtained return. As we chose to work under a supervised learning task and not a reinforced learning one, we use it as a los function instead of a reward.

Features are created in file data_processing.py. The best 98 where selected, and are very diverse: asset-specific values, lags, analysis of the japanese market, time features, etc. The model chosen is based on a specific formatted matrix, with shape i × m × h for i assets, with their m features, and a history of the past h days. Data is then heavily noised using gaussian random noise, in order to prepare the model for "real-world" conditions.

A custom model, which can be found in traders.py, is a neural network with the following layers:

• for each asset i, represented by its m × h matrix, if fed to a dense NN which outputs an array of size i;
• then the outputs of each asset-specific NN are concatenated;
• this array is fed to an i+1 dense NN, with the +1 representing the cash bias;
• finally a softmax activation function is applied to ensure that the output is a unit vector, summing up to 1.

We use the achieved return of the portfolio on the next day, times -1, as a loss function. However, this loss exhibited poor behavior it terms of risk: it almost always chose a single asset. To balance this, we add as a penalizing term the categorical crossentropy between the proposed portfolio and the optimal one. Losses are implemented in traders.py as well.

• interpreting the task as a regression task, and trying to predict the price of each asset for next day: attempts with NN and LightGBM: explore traders.py.
• using Generative Adversarial Network (GAN) to create more training example: see gan.py.
• analysing the press around each asset to predict price, either as such (BERT model trained to predict the price) or as a feature. Explore data_processing.py.

The overall conclusion is that it is able to beat a benchmark (a portfolio that is evenly distributed among each stock), but not the best stock. Moreover, it does not perform well under "real-world" tests, as it is very sensitive to inaccurate/noisy market values measurements. Yes there is still lots of room for improvement!

Rights are granted to anyone under a standard MIT license without any limitation (which means you can do whatever you want with it). Yet I'd love to know if this repo was of any use to you, so don't hesitate to reach out!