The accidental HFT firm | Hacker News Reader

The accidental HFT firm

meanderful.blogspot.com

132 pointsprofquail32 comments

>Things got even better for us in Korea. Now that I had the bits streaming off one by one, or at least byte by byte, you could see that the packets took a long time on the wire. Milliseconds. A common technique in network monitoring is slicing, where you just take the first part of the packet. I did this with the HDLC bypass code and we were now a long way ahead of the game as the bid and ask fields were at the front of the quote packets. Save 100 bytes at 128 kbps and you’re 6.1 milliseconds ahead. As specifications changed and KRX feeds sped up, our advantage reduced but life remained good. Now we were sending out orders before the data packet arrived.

I love it when HFT seems to violate causality.

From the article: “Reification just becomes legacy without a sufficient lack of architecture.”. This is so eloquently put. Technical Debt is a mess in of itself. But there still exists the chasm Of meeting business needs immediately (time to market) and avoiding technical debt. The sign of a good developer is attempting to look ahead and wonder how else will data need to be presented and manipulated - building the system open-ended enough to allow for the evolution of business needs.

> The sign of a good developer is attempting to look ahead and wonder how else will data need to be presented and manipulated - building the system open-ended enough to allow for the evolution of business needs.

I've gradually become convinced of just the opposite: a good developer doesn't think ahead at all. The best way to prepare for the evolution of business needs isn't to add extension hooks in the middle of your code, it's to keep the code as simple as possible so that you can change it in a very direct way as and when you need to.

I agree in the sense that your system will be the most flexible provided you use composeable and easy to understand code as much as possible. Overengineering is overengineering because it trades complexity for "flexibility" but the real goal is to keep complexity down as much as possible.

I've convinced that 95% if the "extension points" I've seen built into code over the years never get used to create extensions - usually because they have significant assumptions about what extensions would be required that turn out to be very wrong.

There are subtle ways to add extension points that most people won’t even notice.

For instance, putting related code into a data structure allows you to add more to it later without replumbing the whole system (as opposed to functions with five arguments).

With enough experience with refactoring, just choosing a structure that’s easy to refactor can serve as an extension point. (I have even tricked people into implementing a feature the way I had it in my head by writing the intermediate version in such a way to suggest a simple change for part 2.)

Never add any extension points you don't already need, but design as if you could suddenly be tasked with adding all kinds of weird ones, but mostly without spending any significant time.

Try to make sure that reasonable extensions are cheap to add, and that even some of the more unreasonable ones are at least possible without complete system redesign.

This should not significantly impact either implementation time, or design time. But it will almost certainly make you avoid some really bad design choices that make quite plausible changes extremely expensive.

Good design is more about avoiding obvious traps than finding a path to nirvana.

It's not always about a hook, sometimes it's about where yo break into subsystems and subfunctions. Breaking it up at the right point allows for easy extension through writing new components later, while requiring the least amount of modifications to the production code (and thus hopefully less bugs introduced into working code-paths). Additionally, how you represent your data may allow for easier extension of the data type or additional, related data type processing later.

I do a lot of data mining, and the first step is always securing the data source and archiving it. There have been times where I wasn't able to revisit the project and finish the actual parser of the data for 5-6 months, at which point I can run through the old data. Other times, I'll write the parser but we'll only be parsing a subset of the data we acqiured. The additional unparsed portionisn't necessarily without merit, it's just not something we are dealing with yet (either because it's still of an unknown quality, or requires more effort than we want to spend at that moment). Knowing it's there and that it may need to be parsed in the future can subtly affect your design such that you leave yourself the openings you'll later need to easily parse this additional data. Sometimes that's as simple as just avoiding situations that would make it extremely cumbersome later, and not necessarily actively making it easier.

I'm not sure ignoring something as obvious as this makes one a good developer, but what's "obvious" and what isn't is actually quite different at different software engineering experience levels.

Depends on which meaure of good in use obviously, but I have never seen anyone capable of being efficient over significant time spans without in some way having design and architecture be guided by principles that balance cost of change, cost of implementation, and difficulty of understanding.

Ignoring cost of change can be absolutely devastating in the long run, but believing you can somehow scry the future for all possible extension points necessary is equally so.

Designing for (anticipated) change in my mind is not at all about adding hooks, API's, or any code at all for that matter.

I think of it more like making sure your system respect analogues to gravity and the other natural laws as best translated into the more abstract relations relevant for the particular kind of system you are building.

One could call it programmer and refactoring friendly design and not miss the target by much.

I don't know if it's because it's early and my coffee isn't done yet, or that's too dense of a phrase, it's just me, or some combination thereof, but it took two reads of that phrase, a complete and thoughtful read of your comment, and then two additional reads of that phrase before I was able to unpack it enough to makes sense of it and how your comment tied into it.

I'm leaning towards the coffee explanation, but multiple negative (or negative connotating) words probably didn't help...

Yeah, it's a bit dense. Sorry about that.

No need to apologize, I agree with the assessment.

In the clear light of day, that comment was just me vomiting words onto the forum in the vain attempt to justify the effort I put into grokking that comment. :/

"The best architecture is no architecture" ... became a popular catchphrase for me in the early nineties. I built a streaming kind of trading system back then that integrated data flow and functional views, as some problems were easier to grok in one and harder in the other. Just thinking about things as simple functions: a = f(b) and code as a way of organising like with like worked for me.

Sometimes the object approach gets a bit silly. Organisation by the first parameter being special doesn't always work too well. For example, GAMS is a better classification for math than simply putting all vector or matrix code in the one classification. Curation is hard at the end of the day.

How much automated/acceptance/regression testing did you guys have?

Would "releases" cause major brick-shitting (like the instance where Mr. L deployed a problematic release)?

It improved from then. We evolved unit tests and full system tests. A release would have to undergo a full simulation to show it was not only problem free but that it still had profitable results that look correct. We had test/debug shims for glibc functions that may cause syscalls, such as time or direct/indirect locale (such as from a printf), that would error on use in code to prevent sneaky costly timing things jumping in.

We did have one rather insidious error where the risk was too strong and preventing some orders going out. It took us a few months to track it down as it was making the system feel clumsy but it was still working. We assumed the market had changed a bit. Cost us a few million and it was a mistake from our best software dev. It didn't change the fact he was our best software dev (he's at Google now). Lesson learnt was to check for the positives and not just the errors.

We introduced a number of risk check evolutions. A separately developed shim was the final check on orders to make sure that regardless of risk the order made sense, e.g. not a zero price. This code was structured to be as independent as possible from code in the main system. That saved our butt a few times. Adding timing throttles was also important to prevent the system reacting in a way that would send a silly number of orders per sec. We also evolved to giving the broker html web pages where they could view the risk in real time and control risk if necessary. This often included integration with broker risk systems and taking their risk files and integrating the constraints to our engines.

At a later firm, we had automated test systems that also ran performance tests with hardware like that which Metamako now provides. The unit tests on code check-in would run not only the unit tests but there was a suite of performance tests where each would reconfigure the network and run things with external performance measurement on the network. This would allow us to track performance bumps of tens of nanoseconds to specific code deltas. Very useful indeed. A slightly customised version of graphite allowed us to see the chart of the performance of all components and tests over time.

Further to this, we evolved to a specific kernel, OS, and bios settings being replicated so that we could reproduce exactly a production system and vice versa. Tuned BIOS and Linux kernels became important.

The risk controls, unit and systems tests are probably the most important things an HFT does. YOLO is very true.

Incredible--was everything written with C++? What was the process like for rolling a new release out to production trading?

The critical stuff was VHDL & C++. Mainly C++. Python & perl around the edges for housekeeping. Can't underestimate the role of bash as there tends to be a lot of scripting of simple components.

Production was getting a branch to pass unit and system tests locally. Then run in an acceptance test environment against the official test exchange. Test exchange wasn't always totally realistic. You'd typically have to mirror some captured production traffic to make it somewhat realistic. Also, some exchanges had slightly different production versus test versions. Trapped us once with some spaces being insignificant in Canada in the spec and the official test exchange but not allowed in production. Uggh. No real way to test for that.

It evolved to linux repo deployment where a yum command would summon the versions & scripts. Convenient for rollbacks too.

Another aspect was testing the ML parameter set. This would typically be updated daily and even though it was not a code change it is like one as it affects behaviour. ML parameter sets would have to pass profit simulation benchmarks to make it to production which was often a challenge in construction and testing on the grid to meet deadlines.

Thank you, thank you, thank you for this. Such an interesting look into the world of HFT. I particularly like the parte of tapping into the cable directly, looking at the first bytes of the packet, and sending the order before the entire packet arrived. Awesome stuff!

Best post I've seen on HN in a month or two -- thanks for posting.

This is great, I wonder what is such a talented developer doing these years? Has tech moved so much that there are less and less HFT jobs? Or is it the opposite?

On a side note - can anyone recommend a book with similar stories? Ie. from the trenches of development, finance + software?

Yes HFT is dying, just check out how many HFT firms either closed down or sold themselves in recent year.

You might want to read book "Flash Boys" by Michael Lewis.

Why is that? Did someone won and took it all ? What happened?

I think it has become more polarised as jitter (as a consequence of latency) has come down. Also more democratised and many vendors provide good solutions.

Still, quite good profits in the system. See: https://meanderful.blogspot.com.au/2018/01/australian-popula...

Could you explain the information edge that asset managers have? I always thought the big guys were the easiest targets.

What kind of algos (you mention that in the article) can they employ to defend or just perform better? I'd love to read more about it, it's very interesting - can you point me towards some resources ? I'm not in finance but this "war" has always fascinated me.

The primary information edge an asset manager has is that they have large enough order that it is not immediately consumable. This will have a price impact. So, if they take the HFT price on offer at a market, the HFT will lose money as the price will go against them.

Most of the algos I'm thinking about there are designed to minimise market impact. They do this by spreading the order over time and/or space (venues) as well as by balancing passivity and aggressiveness.

There is not a terrible lot that is written that is unbiased as most have a particular POV. Market microstructure textbooks are perhaps the healthiest place to start ;-)

Thanks a lot for sharing those bits of knowledge! I'll look into that, have you ever considered writing a book? ;) I added your blog to my long must read!

I love HN for such moments, out of a blue a post on a superinteresting topic AND a knowledgeable person who doesn't mind answering some questions.

Id recommend "Dark Pools" and "The Quants" both by Scott Patterson.

"Flash Boys" is a fun read since its written by Michael Lewis(good storyteller) but Ive heard have claim it isnt very accurate.

Not so talented, just persistent I think. One foot in front of the other helps the journey.

I've taken a year or so off. My aged mother had cancer and subsequent health issues. Doodling with thinking about HFT again, working on some transport IP stuff...

x.25 there is a blast from the past

I was reading through this, seeing some similarities to my work. Then i got a real shock at the end:

> Previously published under the pseudonym Thomas Anderson.

Er, that's me!

I think it's a matrix reference.

"Tell me Mr Anderson, how will you make a phone call... if you're unable to speak?"

That's correct. Automated Trader requested it to be published under a pseudonym when it when out in 2016. I didn't object and chose the Matrix reference.

Reply on news.ycombinator.com