Lessons learned building an ML trading system

The kind of data augmentation I do is adding different candles sizes. I validate with 5m candles, but I train with 2,3,4,5,6,7m ones. I also sample more frequently more recent data. I train jointly with ~22 symbols, but in each X with those symbols, I randomly set some to 0, some I invert their price, some I invert time-wise. This helps generalization for some reason. I tried many kinds of noise, but what I described above is what I found to work best in my case.

I have a more ambitious idea to generate synthetic data using self play: have a bunch of agents trading one against another. This create new price data I can train the agents with, and repeat (this self-play training scheme would be similar to what DeepMind did with AlphaGo/AlphaZero). The issue with it is the need to tune the parameters exactly so that the resulting synthetic data is realistic enough that I can tranfer the agents to real data.

For example, during self-play, should you have only trading agents or should you add "retail traders" that buy during bubbles, "normal buyers" that buy only below, sell above certain prices, institutional buyers that randomly move the price a lot in a given direction. This is a lot of parameters to get right, and it's an optitization problem on it own. You could treat this a as two-fold optimization problem such as in this paper: https://arxiv.org/pdf/1810.02513.pdf, but it gets tricky very fast.

Regulated financial exchanges basically do this already.

Scalability and profitability are orthogonal. If it could scale indefinitely, you'd be right. But no trading strategy can scale indefinitely.

That doesn't say anything about whether or it "works", and it's not a reason to be suspicious of the results, in of itself. All successful trading strategies are capacity constrained.

- The only part I didn't like in your article was how you described creating indicators as exploitation. The limit order book is public by design so all traders can look at it. People have the free choice to trade on a centralized exchange or not. This is a trade-off between revealing information and being able to trade quickly without calling all your friends asking if they want to buy some Bitcoin.

- I'm guessing you used data from other exchanges outside the one you were trading as indicators too. That's unquestionably good since your trading helped information propagate faster or more accurately than it would have otherwise.

- Markets are only zero-sum in isolation. Most participants derive utility from things outside short-term profit and loss. Maybe they trade to manage risk, to hedge, to gamble, have a longer time horizon than you, whatever. They just want to trade and get back to their lives. They don't want to waste time squeezing the last fraction of a basis point out of their fills. It's hard to believe, but they actually enjoy getting picked off, run over, paying too much spread, whatever things make you feel bad or indifferent about the service you provide.

I used to get filled making markets on Nasdaq (which pays resting orders a rebate, and charges crossers) when BX (which pays crossers a rebate) was at the same price, and could lay off the trade for an instant profit. The people who traded with me paid for the luxury of saying "fuck it, send it to good ol' Nasdaq." I used to think it was stupid of them, and from the perspective of a prop trader, it was mind numbingly stupid, but they probably had more productive things to do than read every exchange fee schedule or hook up to every small exchange.

- Providing liquidity has nothing to do with resting limit orders vs. crossing the spread. Providing liquidity is about taking risk off the hands of people that don't want it, and moving it across time to someone else who does. If you're market neutral, trade many round trips every day, and end relatively flat, you've played that intermediary role as a liquidity provider regardless of what order types you use.

- Crossing against mispriced orders is doing the world a favor. You're not the bad guy picking them off. If anything, they're the bad guy for holding the market at an incorrect price.

So maybe think of yourself as more of a service provider. Not only will you feel better, but viewing trading through that lens tends to make you a better trader. Strategies truly built around an exploitation mindset are fundamentally unsustainable, since you run out of people to exploit. Providing a service works forever.

FWIW, the rest of what you wrote is almost exactly how the pros do things. If you built this system yourself, you could make far more than 200k at a prop firm. If you're interested, reply with a throwaway and I can refer you to a friend who's still in the business.