Reinforcement Learning Algo Trading

Background

The success was because they created a system with 3 parts (DNNs, RL, and branching).
They had real players come and validate the moves and strategies.
Finally once they had a powerful model, they trained it further on games against itself.
0 sum game

The stock market is volatile and filled with noise from orders and news.
There are basically 3 key players: retail investors, institutional investors, and independent investors.
Strategies include arbitrage, day trading, value trading, swing trading, gambling, and retail trading.
Unlike the moves in a Go game, the player often has little to no influence on the results of the game unless it has either media influence or lots of money. Moves include (buying, selling/writing, shorting, exercising, publishing, commenting). These moves can be performed on most stocks, derivatives, currencys, or media outlets without any explicit rules.
Not a 0 sum game.

##SkyWing

Sentiment analysis on entities (tickers and businesses)
Business and market reading comprehension. Behind each stock or currency is a business/country/approach, so the system needs to understand this underlying before doing things.

Technical indicators and analysis of these are often used.
System should understand that prices are a result of supply, demand, and a bunch of orders.

A strategy is simply a set of rules for moves, on a set of stocks, specific to some schedule.
Rules can be formulated on combinations of input and output streams.
The system can explore smart or random new strategies at any given date.

Having an adversarial network which responds to the players decisions could be useful.

Backtesting, fronttesting.
Simulating market movements based on media channels is something that needs to be determined whether it works or not.
Randomized chances for unexpected events (black swan, slippage, mergers and acquisitions)