Friday, 19 May 2017

Speed-bump 101

I get a bit tired of some of the silly thoughts that go around about speed-bumps. This is mainly due to the disinformation IEX and Michael Lewis spread about their efficacy. This is a simple meander to cover just the basics for those of us who are not rocket scientists. There is nothing new here. It is a bit quick and hasty, so please feedback any obvious errors so I may clean it up and make it a useful speed-bump 101 meandering.

In the context of NMS rules so far, a symmetric speed-bump fundamentally just makes for a slow exchange. Let's have a quick meander at the simplest example of such a view.

Here is a happy client connected to an exchange and they get a 750-microsecond typical latency or round-trip time (RTT):

The fastest exchange in the US in 2009 was Bats. Bats operated, at one stage in 2009, with a typical RTT of 443 microseconds. Ah, those were the good old days of microseconds instead of nanoseconds and picoseconds. The speed in the diagram is therefore not dissimilar to an old exchange of an antique 2009 vintage. For what it's worth, Bats continues innovating and is now a sub 100-microsecond exchange. Michael Lewis complained in Flash Boys, the infamous piece of fiction, that fast HFT guys and gals could beat you to the punch on your trades by shaving some microseconds off their reaction times. Lewis said you didn't have a chance and a speed-bump would level the playing field for you. Let's dig into that thought.

Here is a sophisticated picture of a simple speed-bumped market:

Guess what? It looks just like the other 750-microsecond exchange to the client. It is still a race to get an order to the public point of presence (POP). It's still a race for reacting to news and external market data from other exchanges. You still want to be as close to the exchange or POP as you can to help you compete in the race. You will still need to carefully plan your trading infrastructure so you don't suffer a disadvantage if latencies are important to you. Microseconds still matter as much as they did with the previous model.

If you didn't know this speed-bumped exchange had a speed bump, would you be able to tell the difference between the two?

No. Here, speed-bumps don't matter. You can't even tell if the bump exists.

Exchanges aren't typically as slow as speed-bumped exchanges. Normal exchanges tend to be a little bit faster. You can probably guess why. A fast exchange may have an RTT of 10-25 microseconds today. Here is a sophisticated diagram of a modest exchange that operates with an expected RTT of 50 microseconds:

It is also just as much of a race to react, replenish, hedge, and otherwise nuance your trading. Given your reactions within your own systems are not necessarily dictated by the exchange, you have exactly the same, or as little, need to be swift to suit your trading agenda. Speed-bumps don't change the competitive necessity, or lack thereof, for client speed.

Fundamentally, a symmetric speed-bump just gives you a slow exchange. 

Slow exchanges are dumb

Why are slow exchanges dumb?

A slow exchange makes you suffer more risk with your trading. If you use the 50-microsecond exchange above, you could have received a fill and hedged with something else, or replenished your market making, and have done that more than a dozen times, before the slow 750-microsecond exchange even gave you a response.

An easy way to bring this home to my simple head is to imagine an exchange that takes an hour to respond alongside another exchange that takes a second to respond. 

Non-farm payrolls come out. 

It's a big number.

You want to react: cover some shorts, and buy some longs, in a controlled fashion where you only take on so much risk, or delta, in bite sized pieces. The one-second responses from the faster exchange are cool as you can get fills, send in orders, and rinse-repeat as suits. Sometimes you'll miss or hit with your orders. You'll be able to adjust fairly easily as you go along. All is good.

Now imagine trying to do the same at the exchange that takes an hour to get back to you. It is still a race to fire in your initial orders and get in the queue. You're still competing. Microseconds matter even on the one hour exchange. However, you won't know if your orders have been filled and what risk you may be wearing for a long time. You can guess and assume you've been filled, or send market orders and hope for the best, or any fill. It is a real mess - a risky mess. 

If you are hooked up to both the one second and the one hour exchange, you could use the one-second exchange's market data as a fair value indicator and try shooting orders with fair prices into the hour exchange hoping for the best. It should be obvious to you that the one second based exchange becomes the price leader and the safest place to trade.

This little thought experiment shows quite clearly, I hope, that faster exchanges, all other things being equal, are the natural hubs for liquidity. They are safer, have better risk, and lead to better price discovery.  Instead of doing just one transaction compared to a slow exchange, you can offer prices, get hit, hedge, and offer further prices once again to the market. This makes for a much-improved market. 

The other thing you need to take away from this thought experiment is that human time scales and computer timescales are different. The two newest speed-bumps are both going to be 350 microseconds. It seems fast. It is around a thousand times faster than the blink of an eye. However, this is a bad benchmark. 350 microseconds is a long time in modern computing. It is likely enough time for your phone to execute over a million processor instructions. Yes, your phone. We need to resist over-anthropomorphising such functionality. Today's faster exchanges can do over twenty round trips of orders in the time it would just take to get one answer from a 350-microsecond speed-bumped exchange. 350 microseconds seems fast, but today it is like an hour, even to someone fifty years old like me.

Slow exchanges are dumb. IEX and NYSE American are slow exchanges.

IEX and NYSE American tell other people about your trades before they tell you

Things aren't quite so simple with the approved NMS speed-bumped models. Both of the SEC approved models report everyone's quote and trade information to the centralised reporting of the SIP feeds before the speed-bump. That is, your competitors listening to the SIP may well receive information before you do if you're simply co-located near your speed-bumped exchange's POP expecting a simple life. It ain't so simple. 

The UTP SIP has a latency of around 17 microseconds and the CTA SIP has a latency of around 80 microseconds for quotes and 110 microseconds for trades. Add in the communication time on the links and your salmon coloured competitors in the diagram above will get the market's information before you do. Dutifully waiting for speed-bumps of 350 microseconds is not how to optimally trade for now. Some people may care about this.

Why speed-bumps?

There was no original reason really. IEX just screwed up their thinking. If you read the Flash Boys fiction, you'd see some of the book talked about Brad's experience of playing catch-up to the rest of the industry by stumbling across a known algo and re-implementing it as the algo he called Thor. Thor sent orders to various trade centres with spaced out timings so they would all arrive simultaneously and thus not leak information between venues. Part of the IEX thinking was that if they had a big speed-bump then they'd have more opportunity not to leak information. That made a bit of sense when the SIPs were very slow from a routeing point of view. However, it was kind of pointless as you don't need to be an exchange for doing that. That is the job of a broker.  Some people seem to use IEX today just as a router rather than for any particular exchange need. Seems a bit pointless, expensive, and inflexible, doesn't it?

Today the SIPs are much faster and your information is not protected. The approved speed-bumps are designed to leak. This highlights that this type of speed-bumped architecture is not sensible.

Despite IEX making a song and dance about keeping only simple order types, IEX invented their complex auto-fading Discretionary PEG (DPEG) order type. DPEG has been joined by their newer auto-fading Primary PEG (PPEG) too.

The speed-bump rationale for DPEG and PPEG is a bit more reasonable. I'll attempt an explanation.

Dark Fader

Most people don't want to be the dumb bunny being traded through when the price ticks. Say the price is at $10.01 / $10.02 and ticks down to $9.99 / $10.00. You'll groan out loud if you've just bought at $10.01. Most market makers try and avoid this by trying to predict if the price is going to tick and get out of the way. It's not the worst outcome for an asset manager as you wanted the stock anyway, but you too would rather have a better price and not be the dumb bunny. However, it is life and death for a market maker as your survival depends on earning the spread and not being adversely selected. If you get traded through on the tick all the time, a market maker will quickly go out of business.

So what IEX did, and what NYSE American is copying, is they put a special algo within the exchange that doesn't get delayed and can pull your order for you, or move the price away, if you like. It fades. That is, the exchange gives their algos advantaged market data to step ahead of you. Michael Lewis referred to this as front-running in his novel, which it is not. It is a form of privileged latency gaming. It is a bit nasty for the market maker, broker's algo, or sophisticated asset manager as it subverts their role in the market. A client of the exchange doesn't have the special non-speed-bumped ability to run algos within the exchange, so any innovation they may dream up is disadvantaged. They can't fairly compete and thus innovation is stifled. This is not something you'd expect to be encouraged, but alas, the SEC has mistakenly allowed it.
Non-speed-bumped access by an exchange's algos prevent clients innovating.
An innovation that kills innovation.

It gets a bit more complicated. These special order types are dark orders or non-displayed. You don't know they are there. They don't mess up the quote feed as you can't see them. These dark orders have a priority below displayed order types, so conventional market making works OK, along with the requisite latency games for displayed orders. However, due to the speed-bump, which allows this special algo access for dark orders, the exchange is a slow one and we should now understand that slow exchanges are dumb. That is, from a lit perspective it is simply a standard exchange, just an excruciatingly slow one that also leaks.


What we have here is a Frankenstein exchange hybrid that marries a dark pool and regular exchange: a Franken-pool. It's quite franked up. No one would ever have won SEC approval for a wholly dark pool to become a public exchange and remain dark. A really slow lit exchange is too dumb for words and makes no sense. By marrying the two together, IEX has managed to fool the SEC into approving a dark pool, the Franken-pool, as a public exchange. From a lit perspective, you can ignore the dark part. From the dark pool perspective, the lit part is virtually just one of many external exchanges.

What is the outcome for the Franken-pool so far? Less than twenty percent of the IEX exchange's total handled shares is lit volume. In fact, there is usually more volume routed to other exchanges than the volume that gets executed as lit. As volumes rise, the dark component percentage is also rising.

As a fine purveyor of risk, trader or investor, you quite often really want to trade when things are happening. You know, when the price is moving. It's part of the utility of the whole marketplace thing. It is disturbingly weird that the speed-bump enabled dark fading order types prevent you from trading at those times. They fade rather than trade. This will also make some of the trade stats look artificially too good on IEX as trades don't happen at times of risk. That is really quite strange for a risk-based utility. More alternative facts running around as statistics is all we need. Also, as IEX's slightly dumb dark fader algo generates lots of false positives, you'll often lose priority to other dark non-fading pegs. Well managed mid-point pegs may dominate dark-fading pegs. The whole IEX pile of franked up dark matter is a bit messy.

The SEC and investors have been hoodwinked. 

I feel the SEC needs to step back and think more deeply about the privileged role of licensed exchanges in public markets. What role and importance does price discovery have? What role, if any, should dark orders play? Do you want to gum up the system and make it less efficient with slow exchanges? Should the national market try to be efficient?

The vested interest noise certainly makes policy hard. The pressure applied to the SEC makes their IEX approval error quite understandable even if it is disappointing. Stasis can be a bad thing. To an extent, mistakes should be expected and their likelihoods even loosely encouraged to hasten the pace of reform, but only if mistakes, like IEX, can be rolled back.

Yeah, I'm not fond of a significant speed-bump as part of a public exchange. Outside the public markets, a parasitic dark-pool as an ATS with a speed-bump that prevents customer innovation may make sense to a limited degree if you can get safe passive inexpensive matches done without leaking information. Just be careful what you wish for.

Happy trading,



  1. What is your view on randomised speedbumps?

    What you have not addressed here is the tax on end-consumers that top-tier speed necessitates and whether that is proportional to the benefits of increased speed.

    If Market Maker A spends $11 million on a network that is 5 micros quicker (but, say, $10 million more expensive than commoditised networks) all other market makers are effectively obligated to also do so. In the end, price discovery occurs 5 micros quicker but all that operational cost is passed onto end-users since market makers are not charities.

    In the end it means network engineers indirectly tax end-consumers of liquidity! A small randomised speedbump avoids the necessity for such games.

    It comes down to proportionality. You used an example of 1s vs 1hour and clearly at such an extreme the benefits outweigh the cost. At microsecond level? Not for end users whose risk management/exposures are far, far longer run.

    1. My simple short view is that it all becomes a game in some regard. A randomised speed bump is not so new. EBS introduced such a beast in the 90s for FX. The simple game is to have more order paths to reduce the effective jitter. It too has a cost.

      Similarly with the idea of short batch auctions, it becomes a latency game of maximising information to be as close to the dead-line as possible.

      IEX have possibly made the worst possible decision by baking in disadvantage to their customers. They would have been better off spending their money on RF to bring the other markets to their customers as well as investing in, or providing, outsourced co-lo which is an excellent and fair leveler.

      An alternative may be exchanged owned gateways on customer sites with guarantee dead-line latencies. Such a thing would be especially useful for fairness in a common clearing market for swaps and other OTCs perhaps. It would certainly make sense for news dissemination too.

      However you slow things down, there becomes a benefit in an alternative faster marketplace do to the intermediation, liquidity, replenish, hedge, risk benefit. Price leadership naturally fits with the faster service. Making a slow service just bakes in a competitive disadvantage that is an economic pressure to overcome but perhaps such a disadvantage is not insurmountable.

  2. Thanks for writing this. I don't understand the speed bump as well as you do, and I'm not sure whether I'm convinced or not. Why is your 1-hour analogy valid? Can you explain the speed-bump mechanics in more detail so its applicability is clearer?

    Thanks for doing this.