TECH TUESDAY is a weekly content series covering all aspects of capital markets technology. TECH TUESDAY is produced in collaboration with Nasdaq.
As we covered a of couple weeks ago, market modernization has moved financial markets from relying on humans and trading pits to trading on optic fiber cables, microwave signals, and lasers. We also showed that microwave and laser are significantly faster than optic fiber (thanks to physics).
Either way, things move faster now. However, one thing that hasn’t changed is that it still takes time for information (and trades) to move between markets, even though that should happen in under a millisecond.
Today, we look at actual action and reaction times between different markets. We see evidence that optic fiber and radio wave dominate in different scenarios. We also see that the price-setting exchange has the lowest latency and, therefore, the lower opportunity for microwave to provide advantages to traders.
How long should messages take to travel around the U.S. stock markets?
First, let’s recap how long messages should take to travel around the market.
As the map below shows, the major exchanges and dark pools are all in New Jersey.
- The New York Stock Exchange and the SIPs it operates (Tape A and B) are in Mahwah in the north.
- Nasdaq and the SIP it operates (Tape C) are in Carteret in the south.
- Most of the other exchanges, as well as brokers and their dark pools, are roughly in the middle, in Secaucus.
Chart 1: Distances between different U.S. trading centers (in miles and microseconds)
We show the theoretical times for microwave and laser (green) and fiber (pink) along with the distances (black).
We see that although optic fiber might be more reliable and have more bandwidth, it is quite a bit slower than microwave. That’s just physics.
In the real world, it also takes time to process and retransmit data – so actual times are likely a little slower. Although some advertised latencies are very close to the theoretical speed of light:
Table 1: Theoretical to advertised fiber and millimeter wave offerings
Dark trades print to the SIP; exchange quotes update in less than 400us
Let’s first look at a recent academic study that found dark trades actually affect lit quotes.
By definition, the market doesn’t know about any dark trades until they “print” on the consolidated tape, or SIP. So, the study looked at how long it took for Exchange quotes to react to dark trades printing to the SIP.
Remember that the SIP is a “golden source” quote, meaning it is published in only one place (although that might change). For Nasdaq listings (Tape C), the SIP is calculated in Carteret. That makes sense because Carteret also usually has the best quotes in Nasdaq listings, so the majority of updates will have the least distance to cover.
Therefore, if we look at just dark trades in a Nasdaq listing (Tape C), we would expect to see that they should print at Nasdaq and then travel out to CBOE first and then to NYSE (at the speeds we show below).
Chart 2: How would other markets react to TRF Trades on Nasdaq listings (Tape C)
SOURCE: GOOGLE MAPS, NASDAQ ECONOMIC RESEARCH
Chart 3 below shows what this paper found. The research found clear “humps” in quote activity that occurred after the SIP published the trade. Interestingly, the delay appears to reflect messages traveling to each exchange on fiber.
For example, the humps below are consistent with the arrows in Chart 2 above:
- As expected, quotes on the Nasdaq update almost immediately. That limits the latency created for order books trading Nasdaq stocks and co-located at Carteret. The delay from SIP time to Exchange reaction time reconciles pretty closely with the processing times published by the UTP (Tape C) SIP.
- The next hump in quote response rates happens at BATS (Secaucus). That’s expected because BATS is closer to the SIP than NYSE.
- Finally, we see a spike in quote activity at NYSE, which is slightly farther away than BATS.
Chart 3: Exchanges quotes spike after TRF trades
It’s important to note that most TRF trades happen in dark pools or wholesalers based in Secaucus. So, those trades also happen sometime before they get to the SIP.
In fact, the delay from the trade time to the exchange reaction time is likely more than double what the chart above shows. During that time, the dark buyer and seller may have already received fills, too, meaning they know their trades are done well before the market knows they have happened.
Trades seem to track at almost the speed of light
We decided to use a similar method to track exchange reactions to lit (on-exchange) trades.
Again, we focus on Nasdaq listings (Tape C). Knowing that the Nasdaq is most frequently setting the NBBO in Nasdaq listings, we should expect to see that a majority of sweep orders include shares sent to Carteret. We should also see consistent and small delays for those trades to print to the SIP, making the main “hump” of the intermarket sweep order (ISO) easier to see.
- We first find a Nasdaq trade in a Nasdaq-listed stock on the Tape C SIP (we make the SIP time = zero).
- We then look for activity in any venue in that ticker for around a millisecond before and after the SIP timestamp.
- We show the activity times using “venue time stamps” (which are included in trades reported to the SIP).
One set of trades we should expect to see is an algo that needs to cross NBBO spreads — using ISO routes — to lift the far touch. Knowing that most brokers and algorithms are located at Secaucus, we should see activity spread across the market starting in Secaucus, as we show in Chart 4 (pink circles reflect bullet numbers):
- It starts by hitting nearby Secaucus exchanges (CBOE, MEMX and MIAX) pretty quickly.
- It then reaches Nasdaq (as it is closer) before printing to the SIP.
- Finally, it hits NYSE (which is fractionally further away from Secaucus).
Chart 4: How would an algo sweep travel in Nasdaq listings (Tape C)
SOURCE: GOOGLE MAPS, NASDAQ ECONOMIC RESEARCH
We show the results in Chart 5 below.
We see a number of activity “humps,” indicating a clustering of trades does, in fact, happen. We also see some of that happen before the Nasdaq trades occur and many more after the SIP updates, although it may be faster than the time it takes the SIP to send data to those venues.
What we expected to see in Chart 4 does seem to occur, and mostly at fiber speeds:
- First, we see a spike in Secaucus trades.
- After roughly the time it takes fiber to get to Carteret, we see a spike in Nasdaq trades (in Carteret).
- Finally, after the Nasdaq trades report to the SIP (at time 0us), we see trades print at NYSE (in Mahwah). Again, the delay roughly equals the speed of fiber from Secaucus to Mahwah.
Importantly, this says nothing about trades at Secaucus and Mahwah making their own way to the SIP. It is quite possible that at this time (step 3 above), the SIP is still reflecting old prices and volumes from both venues.
Chart 5: Watching trading activity spikes in Nasdaq listings, where SIP prints are time zero
Interestingly, we see even more trading humps after the Nasdaq SIP updates. Specifically (circles in Chart 5 for each number):
4. Traders co-located at Nasdaq seem to react quickly to trade dissemination, with a smaller wave of trading starting just after trades print on the SIP (note that this delay in Nasdaq humps matches the processing time published by the UTP (Tape C) SIP).
5. Then, Secaucus exchanges seem to react to the Nasdaq trade with an additional wave of trading, apparently at microwave speed.
6. NYSE also seems to react to the Nasdaq trade with even more trading, also at microwave speed.
7. Finally, IEX sees their activity spike, roughly 350µs (equal to their speed bump time) after the Cboe reaction wave calms.
It’s also worth noting that because most exchanges are not on the NBBO for Nasdaq listings all the time, their proportion of ISO sweeps is much lower (their humps are noticeably lower).
Almost all trades seem to happen in 1ms
Initially, we limited our time window to 1ms around the Nasdaq trade. But in fact, what we saw was the “humps” only seemed to be noticeable in the 250µs before through to 600µs after a Nasdaq trade. After that, the market in that ticker seems to calm to what looks like “background noise,” with activity levels less than 0.2% of the peak levels.
However, this short time window accounts for around 28% of the liquidity in the day. We recently discussed academic research that shows the contribution to price discovery in that window is likely even higher.
Physics makes co-location important
The reason this is important is it shows that if price setting happens at different times and in different venues. It also shows that venues farther away will have “stale” prices for longer – even if they have close to zero processing latency (something even more important to Europe).
This is also especially important given some recent findings that found high-frequency trading tends to benefit when trading against pegged quotes in dark pools located far from price discovery exchanges.
One way to fix that would be to co-locate in Carteret (at least for Nasdaq listings), then trading systems would see prices update almost immediately for the majority of activity in the market.
We’re not saying all fragmentation is bad, but the data suggests that distance between venues makes liquidity harder to source on sweeps and adds to the costs of trading in fragmented markets, even if you’re not trading on an exchange.
Phil Mackintosh is Chief Economist at Nasdaq.
Nicole Torskiy, Economic Research Senior Specialist, contributed to this article.
