i've been trading crypto in large volumes at high frequencies for quite some time now. my models were plain as yogurt feed-forward neural nets. i would engineer some dumb features, sample the at random data, assign the labels (that translate into trading decisions), and train the model. then push to prod, sit back, and relax while the balance grows like a mushroom cloud. just kidding, before that i would grow gray hair while backtesting, debugging issues, etc.
one of the hard problems was labeling the data. knowing that the price is going up 10 bps one minute from now, should i buy? maybe. but what if it's going to crash 100 bps right after this? probably should sell instead.
reinforcement learning promises to eliminate the need to assign labels in the training data. the agent will try a bunch of different variants at random and eventually will choose the most optimal one knowing the state of the working, i.e. the state of the markets. at training time i only need to feed it the features data. another benefit is that backtesting and model training is sort of fused into a single process. rl model is optimizing pnl, and not the label classification score (as in the nn model). with proper train-test-validation split, the most performant rl model can go straight into production (helping me to keep some of my hair brown)
while all the bits and pieces seem straightforward i never managed to tune rl model to work better in the backtest compared to the good old old nn models. maybe i have never been closer to the gold vein, but for now, i abandoned my efforts to build a performant rl agent if favor of nn models.
I made some good money (millions) in 2017 by algo trading crypto. Now in 2018, the bear market is on, but my pnl is still decent. I collected data, trained models, wrote execution strategies, automated everything. I was successful because I was moving fast, trying things, breaking things, etc.
While crypto was and still is my turf, I think I could also do well in the stock market. The problem is that the entry barriers in the stock market are quite large. By my estimates, it will cost between 10k and 100k a month to run an HFT strategy fast enough to compete with the fastest players in the field (e.g. running market making strategies). If you have good alpha you could probably get away with slower and cheaper access.
In the crypto world, the market access is free for all, and everybody has equal standing (from what I know). After implementing some WebSocket/JSON APIs you get access to the market that turns around 100s of millions USD per day. Of course, it's much smaller than the stock market, but it's real nevertheless.
I think it's a myth that smaller strategies cannot compete with established HFT firms. It's also a myth that you need grad level knowledge of math/statistics. It might even hurt, becuase phds will be prone to "do things the right way" as opposed to "do things that work". I think it's also a myth that HFT firms hire exceptional talent. In fact, most firms have rather mediocre staff. The reason is that most firms don't make exceptional money. So they don't have a salary pool large enough to pay exceptional people exceptional wages. I'm talking upward from 300k. Companies like Google will happily pay skilled engineers around that watermark. The Google/startup jobs might feel dull, but the finance jobs are way more boring and less rewarding... I can rant on this forever - lol.
That being said, I consider myself mediocre developer as well. Couple months ago I applied for Senior Developer jobs at 3 firms and didn't get a single job offer. I didn't try hard, didn't prepare for the interviews, but still.
Edit: I applied for these jobs just to see what's up. Neve intended to take the jobs.
I suspect hard work and smarts. I've made money in sports betting and it's mostly grinding through looking opportunities and avoiding bad bets. The smarts part is avoiding bad bets. If you don't know who the sucker is, you're the sucker.
There is so much volume and volatility in Bitcoin markets that a whale can sell insane amount of coins (e.g. 10k) in within days without much effect on price. Placing limit orders in front of the best ask and moving them in case the price goes down will do the trick. Bitfinex traded over 100k coins within last 24 hours. it's reasonable to assume that you can capture 1% of this volume on the daily basis using the limit orders. Selling 10k Bitcoins can be easily done within 10 days without pushing the price down. I think it's possible to do this within 1 day.
one of the hard problems was labeling the data. knowing that the price is going up 10 bps one minute from now, should i buy? maybe. but what if it's going to crash 100 bps right after this? probably should sell instead.
reinforcement learning promises to eliminate the need to assign labels in the training data. the agent will try a bunch of different variants at random and eventually will choose the most optimal one knowing the state of the working, i.e. the state of the markets. at training time i only need to feed it the features data. another benefit is that backtesting and model training is sort of fused into a single process. rl model is optimizing pnl, and not the label classification score (as in the nn model). with proper train-test-validation split, the most performant rl model can go straight into production (helping me to keep some of my hair brown)
while all the bits and pieces seem straightforward i never managed to tune rl model to work better in the backtest compared to the good old old nn models. maybe i have never been closer to the gold vein, but for now, i abandoned my efforts to build a performant rl agent if favor of nn models.
amen.