AI Stock Trading Bot

01/2021 ~ 04/2021

The world is in the craze for stocks right now.

As an engineer, I wondered whether an engineering approach can solve the stock market.

Since the stock market is very complicated, A.I. can be a good candidate for solving the market.
However, the stock market is also highly volatile, so even A.I. would not be able to predict the market price precisely. 

Therefore, I designed A.I. that recommends 4 tickers that have the highest probability to rise. 
For good performances, structure and training data are the most important factors of all.
Below is a structure of my A.I.

Figure 1. The structure of A.I.

The purposes of design are listed below:

• Multihead attention was used to analyze the stock chart in various perspectives. (each head was considered as a trader)
• Analysis of each head was combined with weight vector and FC layers. (a profitable trader's opinion is more important than thoughts of others)
• Combined analysis for the stock chart was further analyzed by BiLSTM and FC layers to score each ticker.
• The A.I recommends top 4 score tickers.

Low-quality train data, which does not have any pattern in it, produce malfunctioning A.I.
Good train data is required for high performance.
Therefore, I only used the top 100 volume tickers data for training.

Specific conditions for training are listed below:

• Epoch: 1024
• Batch size: 64
• Learning rate: 0.001
• Training : Validation : Test = 6 : 2 : 2

After training successfully, I tested the model with actual trading.
However, I could not buy and sell stock manually since I had a hectic research schedule. 
So I developed an auto-trading program with CreonPlus API and Python.
This program automatically trades stocks and reports recommended tickers and profits by Slack.
 

Figure 2. Slack messages from the auto-trading bot

After 27 days of trading, I got about a 3.7% profit margin with this A.I. trading bot.