Automated High-Frequency Trading Bots (GPU, CPU): Live Results Comparison [Institutional Scale]

ElysianSignals

ElysianSignals

Dear Esteemed Members,

I'm asking your discussion regards a community project I planned.

So many people keep asking me about the daily results of different automated High-Frequency Trading (HFT) bots that I could post a comparison regularly in one thread.

I could handle your questions easily if I could refer everyone to this single thread where everyone would find all the relevant daily profits (or losses) from day to day to the bot they're interested in.

People keep asking me because my institutions use both GPU and CPU bots on a large scale. I also understand that most people only want to try one or two and need more information to choose. Clearly, you hold different expectations and can afford uneven resources, while some bots are simply scams.

I don't worry about our edge because everyone who has ever succeeded with an HFT bot knows how much depends on factors like maintenance or server proximity, which competitors can't easily copy without paying the price, which they usually don't.

I emphasize that it's my personal contribution without the involvement of my institutions because some bots my institutions use I bought for myself. I only wanted to explain they can scale up to the institutional level. Consequentially, I'll share the results from my servers, and thus, everything I write is a personal opinion and represents only my success or failure.

I believe the most straightforward method to figure out if you find it useful is to keep sharing daily reports in this thread. Upon demand, I prepared a live stream where people could follow everything in real-time. Clearly, the sustenance of such a stream has costs, and I'll do it only if you like the principle.

Being a personal venture, I won't ask my employees to involve. So, I've to ask you to help me start it off. Let's start with some questions. When you answer these questions, consider what I've written above. I'm looking for a solution I can implement alone.

1. How can we describe results, and in what format?

2. How many bots can we compare with human eyes on the results table?

3. What analytics do we want to see together with the results?

Example answer:

I want to see the daily profit, maximum drawdown, pair, trade amount, open account value, and close account value.

Compare no more than 6 bots on a bar diagram.

It'd be nice to see the daily profit deviation of individual bots compared to the average daily profit of all bots.


Kind Regards,
Ely
 
destriero said:
Are you using MT5?
More...

I use MetaTrader 5 but wouldn't consider MT the best for HFT. I meant platforms like Virtu Apex, B2B Gateway, Aladdin, Talon, and Lattice. I usually employ on-premise tools as well like Python trading scripts.

qlai said:
Hello, could you please first give us an idea which instruments you are trading, from which locations, and average latencies required,
More...

You might have noticed I wrote about servers (not servers) because different bots require particular locations. I figure the NY4 data center in New York City is suitable for trading NYSE equities like ES. I also use NY servers and Chicago locations in CL options' algorithms to connect CME, NYMEX, and CBOE. Trading LIFFE futures, the LD4 data center in London usually gives acceptable results with TYX.

I trade BTC/USDT and ETH/USDT on exchanges like Binance and Coinbase. The deal with these platforms is their tendency to change (or expand?) locations, and they keep forgetting to notify me. So, every time my latency drops, I've to contact them to figure out what happened. Usually, it turns out they only bought new servers, redirected traffic, or moved.

My average HFT latency is around 10 µs. But that's what I could build as an individual. My latency is not the required latency. It's a latency that's acceptable in the competition. You probably could make a profit with slower systems, too.

In your company, you could build faster with proper staff and more resources. I guess somewhere in the single digits. I can confirm it's possible and not unique.

At average, the latencies between individual friends and me range from 10 (µs) to 50 (µs) while 50 (µs) still being good enough to profit from particular HFT strategies.

Slow Learning Elf said:
I hear carrier pigeons have pretty good latency!
More...

True, but in my experience, carrier pigeons have a high percentage of package loss. ;)

BlackPhoenix said:
Why are you using GPUs for HFT?
More...

I use them to analyze large amounts of data and make quick decisions to identify arbitrage opportunities, statistical patterns, and market inefficiencies.

GPUs are much faster than CPUs at performing certain types of computations, such as matrix multiplication and Fourier transforms. In addition, GPUs can be better at parallelizing workloads, which can further improve algorithmic trading.

(Every AI creator knows how important parallelizing is in artificial neural networks, but people usually run away when I mention AI. So, I'll explain its uses outside of machine learning.)

I use GPUs in market making, order book analytics, and calculating order costs to route orders to the exchange or liquidity provider that offers the best price.

Finally, I can build clusters from GPUs that are more efficient than my best personal computers.
 
Interesting, I don't hear much people here talking about using GPUs. I'm currently working on doing algo parameter optimization on GPUs. What platform do you use to utilize GPUs and are you using OpenCL or something else like CUDA?
 
Cool thread.

Ely, can I run HFT from my home in Tahoe? Problem is that local zoning won’t allow me to build a microwave tower above the tree-line. Any suggestions?

What specific bots did you purchase?

thanks.
 
You must log in or register to reply here.
Back
Top