Men, HFt is HFT is not your problem couse they dont care about you, not your war. But if you want a profit, you wont get from HFT couse they hold cents, there is no money to give you men. If you get into where is 0 resilency (ex.news) you will be punched probably like when there is a train smashing a penny.
Read his last posting. You will see he is is not the idiot you think he is. But as always you know everything better. And the language you use is as usual too. Vulgair and used by marginal people. Is it really impossible for you to show at least a minimum of respect for other people? You can disagree and discuss, but what you do has nothing to do with discussions. Have a nice day too.
Really? What does "You can use point specifics ASICs like our Bitcoin folks and get to 100x faster than intel xeones, but the problem with that approach is that any minor tweak (say change a + to a -) in the algorithm will cost millions of dollars + months of wait time... the algos will not be adaptable which is contrary to hft requirements." even mean? Jump Trading and other hft firms adjust algos all the time. Stating that those algorithms cannot simply be adjusted and changed is simply not reflecting the truth. Even FPGA logic can be easily adjusted via a simple code change. Nothing of that costs a penny and the time it takes to validate code changes, testing, and have management sign off completely depends on each firm and even each group. Just because someone tries to sound smart (and maybe he is) does not mean what he said is applicable here nor factually correct.
His intelligence or lack of intelligence is unrelated to whether HFT can afford a few clock cycles to use complex tactics. The time it would consume is 'relative' to the time others are consuming thus it's likely they have both speed and intelligent tactics. It's unlikely dumb tactics alone can be speeded up enough to consistently win in this competitive environment. I fail to see why HFT wouldn't use intelligent tactics.
Read my post again; I never said that HFTs don't/can't change their algos and I never stated that you can't change algos in FPGAs. What I stated is this: if you want to design your own ASICs using highly optimized lookup tables in order to get 850GHz+ (bitcoin folks achieved this), it is highly complex to modify that asics chip and takes a fortune and months of wait time. However, if you don't go with complex optimizations, and stick with general purpose asics/fpgas, your clock frequency will be about 2.6GHz at best for both ASIC & (little lower) for FPGAs. Intel xeon gets clock speeds of about 3.8ghz+ because they use a special silicon die and it is far more superior than the dies that TSMC's/samsung uses for their fpgas/asics... HFTs use FPGAs to bypass the pci bus which adds unnecessary latency... and yes, they do change the firmware often since fpgas are reprogrammable. .... and so my point is, since you are stuck with 2-3.8ghz clock speeds anyway, the amount of code that you put into the critical path so that a system can respond in 1us is limited. For example, try doing a quick sort of about 500 numbers when a tick arrives and see how processing time changes... My point is, you can't put intelligent number crunching AI/neural network/machine learning module and expect the system to respond under 1us... HFT firms employ highly intelligent sw engineers how squeeze optimized code and outsmart each others' systems... but the functionality can't in anywhere be compared with the number crunching financial analysis that something like R (or whatever package) provides... Yes thats true. Its a rat race... Meaning of intelligence is subjective in this context. I consider neural nw/AI as intelligence... and the rest is simple code/logic.
I can't imagine anyone doing this in realtime rather than offline during backtesting. Btw, Volpunter specializes in the area you referred to ..... he is extremely skilled at reducing the time it takes to execute code.
My sincere apologies. I did not notice you talked purely in the context of non standard fpga boards. Most hft shops to my knowledge go as far as standardized fpga (I am sure there are exceptions) and I think the reason is that one has to make a compromise between latency and required overhead to execute logic. Despite the simplicity of some of the market making algorithms there is still logic involved that at times takes into account parsing, sorting and other low level building blocks that cost latency but that are necessary.
Accepted You are right... if one wants to respond within 1us, we have 1000ns/0.2 (4 ghz processor) = ~5000 instructions... of course we can argue about hw optimizations (l1/l2 cache misses etc), ... how much "useful" functionality can one pack within 5000 assembly instructions apart from the operations you have mentioned... its hard... i have more than a decade experience in system sw, and I give up here. Whatever code you put in within that 1us... I don't know how it can outsmart human intelligence... speed does... I agree... anyway, this is how I see HFTs vs humans...
I agree with everything you said. The issue I see with hft is that it is purely about latencies and as soon as regulators shut the door the game will be over. Not just phased out over time but potentially overnight. It was a great game that worked for quite a long time for those willing to make huge technology investments. Equity markets in the US have gotten to a state where I think it cannot get worse in terms of inequitable markets. It is just not a fair game strictly in terms of fair electronic executions. The situation is very different in markets such as Hong Kong or Korea (though in HK in particular one battles other issues such as rampant insider trading)