I actually believe that trading is an interesting problem that should be studied more in Academia and Machine Learning. It has many aspects (sparse rewards, long-time horizons, simulation-to-real-world transfer, non-stationary data distributions, etc) that current ML algorithms struggle with.
Unfortunately it seem like most ML people are not really interested in trading, perhaps because it has such a bad reputation (which is IMO unjustified) - so they work on games instead :)
"most ML people are not really interested in trading"
You couldn't be more wrong on this. Stock market trading has the lowest barrier to entry of any endeavor. All you need is $1000 and Robinhood account, which you can open one on your phone in 5 min or less.
I've been following HN for a while, every time someone comes up with a trading algo or posts a link to algo, there's were hundreds of upvotes, lots of comments.
i think it’s just because they want to publish, and as you say it’s not easy to find a really good publicly available dataset or simulator. I think if these were publicly available and there were a Python package, people would rapidly get interested in RL for trading. (if it even works for trading — i don’t know much about it but maybe simpler techniques work best, in which case there would be little chance of producing a publishable paper.)
Ther are plenty of free datasets out there. You can get upward of 10 yrs of daily OHLC stock data on yahoo finance. The amazing thing is yf has S&P 500 index since 1927. Free!
Quandl has many free, or low cost stock market/commodity datasets.
I'm not sure what you mean by a "simulator". One of the greatest challenge applying RL to stock mkt is precisely that the market itself is not a MDP.
I don't think daily OHLCV data is a good data source. First of all, it's too little because of the data distribution shift over time. It's also driven significantly by outliers and events outside of the data (news, etc). There's way too much noise in daily prices that most of the signal is drowned out (longer time horizons = more uncertainty). I don't believe you can find any edge looking at daily data. This kind of data is would be equivalent to what MNIST is in ML. Nice for some playing around, but nobody who is serious would use it for production or benchmarking, at least not by itself.
There is a good reason trading firms pay a lot of money (sometimes millions) for fine-grained historical data from exchanges. It's not only about speed. For interesting experiments you IMO need L2 or L3 order book data, ideally somewhere on second or sub-second scales. That's not HFT (which is nano and micros), but somewhere in the "middle" - it's a different world than what you are talking about.
By simulators he means market simulators for L2/L3 data with a matching engine, latencies, queue positions, jitter, complex order types, etc. You can't simulate other market participants (at least not fully, but there are techniques to even estimate this based on live trading feedback), but there are still many things left that you can simulate in a realistic way during training and backtesting. Trading companies typically have their own high-performance simulators built in house. Some of these are incredibly complex. Good simulators can give you a huge edge and are absolutely necessary.
What you said about daily data is precisely what makes stock mkt so interesting and challenging : nonstationarity.
"outliers and events outside of the data, news" : these are precisely the stuff your models need to learn, and the fact that you consider them noise tells me most folks have no clue how to predict these "noise".
I would actually say the opposite - its much easier to get hold of a financial dataset than for other fields. There are packages for easily downloading Yahoo finance data e.g https://github.com/ranaroussi/yfinance
Unfortunately it seem like most ML people are not really interested in trading, perhaps because it has such a bad reputation (which is IMO unjustified) - so they work on games instead :)