Lessons learned building an ML trading system that turned $5k into $200k (2019)
tradientblog.com
tradientblog.com
Clicked, read, well -- it's about crypto trading.
Putting the word "crypto" in title might have been better.
A bad strategy could lose 200k over the course of a day if allowed to run, but again I think the poster was referring to discrete events in a short period.
Crazy events tend to be money printers outside of the initial bad trades since there’s a ton of volume, wide spreads, and general chaos following. Most MMs run taker strategies as well, and in fact many of the large ones make most of their money taking even if volume is evenly split.
More realistically, a strat without edge gets eaten alive by costs, or by toxic flow (as other MM are either better at pricing or faster at getting out of the way).
But yes a decent HFT MM will have down and out circuit breakers.
That's not exactly what the research says. It says that fund managers ask for more in fees than they earn you back in capital gains.
(Which is obvious: if they couldn't charge their customers more than they earn, they wouldn't offer their services in the first place -- they would just invest on their own.)
This isn't obviously true. Taking 1% yearly of $100,000,000 (compounding) from investors is better than earning 15% yearly on investing your own $1,000,000 (compounding). (Or some similar charging strategy).
However, from my experience, when looking for profits, people tend to prefer to arb inefficient markets before they choose to work on larger scales. So if one can choose between charging more than one provides, or charging "fairly" but at giant scales, the first step of the evolution will be the arbitrage.
Imagine you knew, 100%, a 20% bounce coming. You may be able to scraggle together a few hundred thousand, maybe a million, and do OK.
Or you can find people with millions or billions in total assets, skim 5% or even 10% off the top, and make out like a bandit.
All of that said, on aggregate, you're probably right.
At evening, one friend is tired and goes sleep, the other one has 100$ bill in the pocket and wants to try his luck.
He goes to a casino and puts 100$ on red and wins 1,000$. He then puts these 1,000$ on red and wins 100,000$. He puts again on red and wins 1,000,000$. Lastly, he puts all on red and loses all.
Later, when he is back to his hotel room, his other friend asks: "how was at the casino?"
To that he replies: "Oh, I just lost 100$".
This is interesting! Standard wisdom says that you get more statistical power from predicting a continuous variable as such, and then applying a threshold on the output, rather than trying to model the classification directly. Modeling the dichotomisation directly is equivalent to throwing a third of the data in the rubbish bin: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2972292/#__sec1...
They therefore have a vested interest in filling the market with more such suboptimal algorithms that they can exploit for profit.
1. You’re connected to more venues to your counterparty, and can hedge out profitably even if their trade was good.
2. The loss is conditional on trading with you (maybe you can price a derivative very well). The counter parties might on average make money, except when you’re sitting on the right side of a mispriced derivative.
3. You lose money when trading with other bots, but are good at competing to provide liquidity to other traders who are not toxic on a short term
The early crypto market was absolutely disgustingly inefficient and easy to succeed as a liquidity provider in. Simply understanding how to price a perpetual would have done much better for them than fancy ML trading.
https://web.archive.org/web/20221008065200/https://www.tradi...
A good overview of the field, and many insights are not crypto-specific. Best sentence in the OP:
> To be profitable, our trades must be good enough to offset all trading costs.
This is advised is even ignored in many academic studies that would otherwise no longer be attractive.
Well... Except if the decisions are game-theory-optimal.