First of all, this has nothing to do with High-Frequency Trading. It's about the NYSE not delivering a product (SIP real-time data) while collecting $100M a year for that service. This was happening for at least 3 years.

I am a champion of free markets. The term "High-Frequency Trading critic" is a label others use when they either can't understand and/or refute solid evidence.

I'm happy to answer questions.

Do you have any info on how often lag was present, or how often over a specific threshold?

Do you know how representative you sample was with respect to when lag was present?

I understand you may not have gone into this much detail in your research, but if you have any info like this, I think it helps illustrate the scope of the problem.

I've also posted many examples on twitter @nanexllc using recent data (yes, this is still a problem!)

I have always been facinated by the underlying technology that drives Nanex's products. I follow http://www.nanex.net/NxResearch/ regularly and would like to see some more technical anotomy of some of the tools you guys have built.

Thanks and great work!

And, just to be clear, the NYSE generously agreed to pay a $5M fine as punishment?

The actual SEC order is linked to by Nanex, and can be found here: http://sec.gov/litigation/admin/2012/34-67857.pdf It explains NYSE's failings in more detail, e.g. "Inadequate Compliance Efforts" and "Failure to Retain Business Records".

Covington and Burling - who represented the NYSE in this matter, used the paltry $5M fine as bragging rights in their year end letter to clients! 'We got the NYSE off with just $5M'

Nice.

I printed one out and put it on my "Tufte" shelf [1].

Edit: Add [1] https://en.wikipedia.org/wiki/Edward_Tufte

At my last job I worked on one of the Reg CAT bids, and I spent a lot of time looking at Nanex Research--totally mind blowing. The markets would be a better place if the regulators were doing Nanex-style analysis. Keep up the good work!

All the while they are selling non-public information as a product that is deliberately used to micro-manipulate the market?

I mean, deep down I always knew this, but to have it all spelled out is shocking because it exposes the NYSE as systemically fraudulent. The implication of that is so overwhelming it causes me to want to ignore it out of sheer inability to think of a way to solve the problem.

This did not affect you as a retail investor, as you get the NBBO price. This only affected you if you were an HFT firm with bad infrastructure who depended on the aggregate feed and not their direct line.

Like tptacek said, the price for getting a direct line, while expensive, is not unreasonable for a business. The largest cost will be salaries for the people writing your code and maintaining your infrastructure.

Maybe I misunderstood, but it seems like they're saying that's not the case:

FTA: "First, we had to rule out the possibility of the delay being caused by the consolidation process (which people often erroneously point out)."

They get their data from NYSE from apparently an aggregate feed, which was unintentionally delayed by NYSE. In my opinion, they should have been getting the direct feed and doing all consolidation themselves.

However, HFT firms will still have a vested interest in making more efficient aggregate feeds than the consolidators/exchanges.

Be that as it may in reality, per below quote I assume that behavior is still illegal.

"A crucial sub-ruling in the regulations prohibits exchanges from giving stock quotes to special groups faster than to the public. This sub-rule is of such importance, that without it, the rest of the rules (Reg. NMS) essentially become meaningless."

Please tell, how the world is supposed to provide that for free to anyone who wants it?

Colocated servers are the public.

The consolidated feeds are necessarily slower than the direct feeds, due to processing delays, unless the exchanges intentionally slow down their direct feeds. There are several companies that will sell you expensive FPGA or IBM Cell-based cards that will locally create a consolidated feed from all of the direct feeds. These cards aren't cheap, but are necessary for trading latency arbitrage strategies, and are also useful for low-latency market making.

Of course the consolidated feed requires a slight processing delay, so unless they intentionally slow down the direct feed, quotes from NYSE will appear first in the direct feed and very slightly later in the SIP feed. The argument here is about if the delay between SIP and the direct feed is small enough to be permissible.

The "public" being talked about is the paying public. Nothing in the regulation requires a free SIP feed, and I assure you NYSE doesn't provide a free real-time SIP feed. The regulation is to prevent selling an intentionally slow connection to everyone, and giving a faster feed to only their top tier customers.

Disclaimer: I'm not a lawyer.

Pretty simple.

The exchanges take in $500M a year in SIP fees, which are supposed to go to the most efficient hardware and software available. Pretty sure that's not happening.

Oversimplification, forgetting the laws of geometry and physics.

The triangle inequality and relativity conspire against the SIP. If you're getting the SIP from B, any new prices at C go to B and then forwarded to you, so it's faster (best case equal in the degenerate triangle case) to get direct feeds from B and C and consolidate locally.

I was just pointing out that

> If a private company can consolidate faster than the SIP, then the SIP isn't using the most efficient hardware and software available now is it?

is faulty reasoning.

But the law's the law. "It doesn't make sense" isn't a valid defense.

Usually due to capex: energy infrastructure, network effect: financial, or limited availability of resources: wireless spectrum.

If it's in the public interest to have trucks and no firewood seller would otherwise provide them. That's absolutely something government can (and IMHO should) mandate in exchange for a firewood license.

The retail investor gets the slowed down price (the NBBO is really what the internalizer/wholesaler saw from the SIP when they executed your order). That same internalizer buys/sells from the faster feed. Whenever the two feeds are out of sync, it's like printing money.

"The price for getting a direct line is not unreasonable?" I just laughed out loud reading that. Is $60,000/month for just one exchange (there are a dozen), reasonable? Then you need network engineers, infrastructure, etc. It used to be nowhere near this expensive for "real-time" data. Orders of magnitude less.

How much latency would be acceptable for a fully consistent, high throughput distributed log of events, that when it becomes unavailable the entirety of US trading shuts down and if it ever gets out of order can cause gigantic lawsuits?

How would you change the obvious organizational disincentives as well? Given that maintenance of the SIP is full cost, no profit operation for the exchanges?

Finally, don't take this as a criticism of your whistle blowing, I think that is critical to a functioning market, I am curious to know how you would solve the obvious technical challenges the consolidated feed faces?

I'd still love to hear you discuss it...

Similarly, people complain about the SIP feeds being latent and I'm sure it would be a very real problem if any single exchange being unavailable made the entirety of the SIP unavailable.

Its a fascinating problem because the engineering trade-offs are relatively static and well known, but each and every decision you could make provides an opportunity for arbitrage market activity that someone would consider predatory.

That we've layered a regulatory requirement that is pretty vague on some of the edge cases on top of this, just adds to it.

If you are asking about evidence that some people have claimed that the SIP being non-consistent is a problem for audit and regulatory compliance, is that not what you yourself are arguing in this post? http://www.nanex.net/Research/IsNBBOIgnored.html

I think maybe you are reading into my comment something that isn't there?

Edit: wow, rate limited after three posts this morning. A new HN low.

My response to tptacek below:

My dumb order?

How can HFT's intercept and redirect my trades to their, appently, captive pool of dumb trades?

Sure sounds like multiple markets are operating...and you are even telling me that my orders will be scraped before they can even reach some markets.

>is that HFT market makers will give you a price break to route your dumb order to them

You claim a price break on a private exchange. How can they (HFT market makers) deliver a better price than the public exchange? Presumably they cannot have bought it on the public market at the better price, so some party must absorb the loss on the price break.

My question can now be translated to:

Who absorbs the loss from the price break?

As a small trader, you get access to special lower prices that a hedge fund can't get. Your broker will be routing your order to a wholesaler who will fill it at lower prices (ie, narrower spreads) than you would get on the open market, because they have a legal obligation to not screw you over.

You can request that they route it to anywhere you want, and they are legally obligated to do it if you ask, but you really don't want to do that. Unless you like giving money away of course.

(Also, you don't seem to understand basic market mechanics. You use words like "scraped" or "picked up" which are nonsensical in context.)

IEX recently suggested that retail investors should request their broker to route their orders to IEX, for which they got criticised very harshly. And rightly so.

I used the word scrape in response to tptacek's response below my above post. If a trader is not aware that what tptacek describes is happening to their orders, then I believe my characterization is accurate.

As a small investor, your order is probably routed to a wholesaler like Citadel, where you will get a better price than you would otherwise get.

And you are upset because you incorrectly believed that your order would be routed somewhere else (where, exactly?), where you would get a worse price, which you think would be better because...why?

> I believe my characterization is accurate.

It's not even slightly accurate; it's literally the opposite of the truth.

Would you trust Wall Street to give you the best price on anything if there was no way to verify it?

There is no way to reasonably know if you are getting the best price (because they can change many times in the same second).

You are relying on the internalizer to give you the best of many possible prices. History says: "that's not going to happen"

Full disclosure, I have in fact worked with your data and at an HFT in the past but have moved out of the industry, if you were going to ask.

1) A dark pool. This is just another name for an exchange, but one that is not public. I don't know if NYSE runs one, but it wouldn't be surprising. The theory behind these is that they are provided as a service to large institutional investors to trade with each others outside the vagaries of the public markets for lots of reasons (mostly advertised as a way to not impact prices). The reality of these has been mixed at best. You have to opt into using a dark pool.

2) Payment for order flow. In this case your broker (not the exchange) has an agreement with a third party to sell your trades to them so that they can either trade with you directly or route the trade to an exchange. The reason they want to do this is that the aggregate of all the trades they are buying is not price opinionated (as opposed to hedge funds say) so the market maker can make the aggregate difference between all of them. This mechanism is largely how all discount brokers offer cheap or free trading, so it is likely a huge boon to you. A couple of things to note about this, your broker has to disclose it to you, I believe you are still required to get the NBBO price, & I believe you are required to be able to opt out by directing your order to be directly routed.

[edit] To clarify your question `jsprogrammer `tptacek is referencing #2 above. The reason the technical oddities & violations at NYSE don't have a huge impact on a retail investor is that unless otherwise stated, they are interacting with a wholesale market maker before their order hits any exchange, in a deal worked out with your broker and governed by legal obligations on both their parts.

A U.S. equities exchange has to be registered with and approved by the SEC. There are currently 12 of them. As a registered exchange you have to play by certain rules, and in exchange your displayed top of of book is subject to the order protection rule, meaning that no other venues (exchanges, dark pools, ECNs, or broker dealers) are allowed to trade through your top of book quote without first routing an order to you for the full amount of the displayed size. Exchanges also get a share of SIP revenue (which is important to exchange economics).

In it's most technical definition, a dark pool is an ATS (specifically not an exchange) that has no displayed quote i.e. no market data whatsoever. Most ATSs in the US fit this definition, though there are also ATSs that display a quote. For example, IEX is an ATS that has feed for their top of book quote... so technically not a "dark" pool.

NYSE doesn't run a dark pool and it would be weird if they did. Most US equity exchanges support fully hidden (completely non-displayed) orders, so if you want to trade in the "dark" on NYSE you can just send a hidden order. I think this is one area where futures exchanges and equity exchanges differ. In futures you have icebergs but not fully hidden.

2)

Wholesalers aren't allowed to give you a worse fill than the NBBO. That's the law. Often they will actually give you a tiny bit better than the NBBO on marketable orders. I think they do this so retail brokers can advertise price improvement. For most of the low cost retail platforms there is no ability to direct your order. You give it to the broker and I think basically if it's a marketable order they send it to the wholesaler and if not they post it to an exchange that offers a liquidity adding rebate (the brokerage keeps the rebate, natch). If you are willing to spend a little more on your brokerage account, you can use a platform like Interactive Brokers which does let you route orders directly.

Thanks for the clarification on the NBBO ramifications on wholesalers. I have never worked on a order flow trade and would love to see an insider account of them (its sort of what I was hoping Flash Boys would be about).

This is the case for retail orders, but not for trading firms. Prop trading firms don't tend to route their orders through wholesalers; they have deals with execution platforms (or implement their own) and route their orders to exchanges.

But there's a whole 'nother set of reasons why we're not especially concerned about competitions between prop trading firms on speed/latency/whatever: if they're impacted by HFT at all, they're competing in a zero-sum game to arb prices, and there's no moral reason why a slower trading firm is entitled to some equitable share of all the available profit from some arbitrage strategy.

If you were a retail investor or a mutual fund, what NYSE was doing probably didn't impact you at all.

   Q (CNBC): Should long-term investors care about milliseconds?

   A (Hunsader): Yes, they should.

Then where does all the money HFT firms make come from?

Second, HFT firms compete with other finance firms, so what money they do make comes from bidding down the costs those firms were imposing on the rest of the market.

If you're a retail trader, automated electronic traders make money off you by outbidding the markets to quote good prices to you for your dumb market orders, and then pocketing the spread --- in other words, they make money the same way market makers have always made money; they just make less of it at an instant, but at a far larger scale.

The same is true of mutual funds, which is why Vanguard --- the firm most people would say is the most trustworthy in all of finance --- has repeatedly claimed that HFT has lowered their cost of trades.

That seems like a lot of cost everyone could save on by simply having the exchange enforce some minimum timings.

There are two straightforward problems with microsecond-speed electronic trading:

* At very small timescales, possibly as a sort of inevitable consequence of the CAP theorem, correlations between instruments that should trade in lock step start to break down. Since you can make money correcting these mispricings, time and energy gets sunk into doing that, and that imposes costs on the rest of the market. It's uncertain how high those costs are, but they are small, and clearly a pittance compared to non-automated market making.

* Electronic trading provides very nice paychecks for pretty interesting technical work, which means some talent gets attracted to the market that would otherwise get paid less money doing something with more social value. How big a deal this is to you depends a lot on your worldview.

To me, neither of these issues seems worth the galactically high cost of restructuring all the markets in such a way that we'd reliably avoid these problems.

I don't mean to get rid of automatic market making, I'm saying the exchange could just move to a system where it checkpoints every 10ms or whatever, still much faster than humans can blink, but no detriment to the algorithms. Then everyone has 10ms to come up with orders that make sense in the current market situation, instead of the ridiculous timing game.

I'm not saying the markets must function this way, only that they do, and it's not evidence of a giant conspiracy or grift that they do.

The rest of my comment suggests that while we could potentially quantize the markets, it doesn't seem like we have much to win from doing that, and, further, that it's likely that the problems we're trying to solve by doing that would persist.

The window does nothing to remove the need for speed.

The world is a continuous system. No matter how hard you try you can't force it to be discrete.

(There are other problems in your proposed scenario to having to do with mismatched buy vs sell volume, but we'll ignore those for now.)

I think the above poster meant to imply that the 10ms delay would cover all market transactions, so if you put in an order now, then no one knows it till the next 10ms stop post.

Essentially its buffering all the data then releasing it at once, so it's not a truly continuous system anymore.

Trades occurring at other locations. Even if the other location are operating in discretized time, there's going to be variability in how fast different market participants can transmit information from one location to another. From Wolfram Alpha, Tokyo to NYC is 36ms at speed of light in vacuum and 51ms at speed of light in fiber, so someone with a faster network could be multiple discretized ticks ahead.

Stuff that happens in the real world. The stuff that all security prices are eventually based on. It's not a closed system you know?

For example, infrequent important events will result in a race for people to submit at the last microsecond, but because they're infrequent 99% of the time the 10ms buffering will effectively change the behaviour of market participants.

You can design the rules to optimize for one thing or another, continuous world notwithstanding. I'm suggesting a way to optimize for something more useful.

Are you perhaps confusing exchange in this case with a broker?

This is good for you, not bad.

What changes? What's the end result? How do normal people or society benefit? What would have to happen for the change to be declared a success. Conversely, what would have to happen for the change to be declared a failure?

What exactly is an "Exotic quote type"?

Do you mean order type because there is a huge difference. If you don't understand the difference it makes it hard to take your assertion seriously.

1) It is a common fallacy that there is "the exchange" or "the market". The market is an aggregation of exchanges (and non-exchange traded instruments) across a wide variety of jurisdictions and localities. But even though there is no central coordination between all that jumble, the instruments are correlated. That is, the change in price, demand, or availability of one is a signal about the pice, demand and availability of the others.

2) The market is made up of participants who exist on a band of price/time sensitivity. That is, someone like a retail trader doesn't have extreme price sensitivity but may have time requirements. In trading that style of participant is known as "dumb" (oh traders, you so PC).

3) Market makers make all their money trading spread for time but thats not where the latency games come into play. The latency is because...

4) Smart money is price sensitive (or price informed) and they tend to be where market makers lose money. Smart money tends to be either prop traders with particular algos, large block hedge funds or other market makers. Those are the people who cause market makers to play latency games, because its a vector a market maker can control that keeps them smart. They will never be smarter than the hedge fund that knows its going to buy a bunch of shares and therefore drive up the price, but they can at least jump out of the way quicker than the other smart money.

So in order to be successful, especially when your margins are razor thin (and they pretty much always are), you have to be very, very good at canceling your orders as quickly as possible any time it looks like the market is likely to move against you.

I.E. how much money is a good HFT firm making at scale?

Actual results for 2015. Your numbers are a fair bit off of those. What you can see from the results is that their revenue/profit numbers are small in real terms, but their margins are big compared to other businesses, even technology ones.

Personally, I'm not offended by that and think that the entire conversation about HFT is funny given how little money is actually involved, but if you view profitability in any context as evil or rent seeking, then you have ammunition for that argument with Virtu.

[edit]

Another good one to look at is Knight (Getco) http://investors.kcg.com/phoenix.zhtml?c=105070&p=irol-SECTe....

Had a much better year last year vs the previous 2 and it still was a fundamentally little amount of money.

[edit again]

That Knight statement is very interesting because it shows that the vast amount of their profit came from execution services and that their market making business was very low margin. Now that could be because their market making business is taking the full shot of the infrastructure accounting that enables the execution services business, but its interesting none the less.

> Full Year 2015:

> Net Income of $197.5 million

> Adjusted Net Trading Income of $500.7 million

His numbers seem pretty accurate.

That said, whenever any trading entity talks about their trading income, be suspicious, because that is usually just the difference between bought/sold (and fees) and doesn't account for the infrastructure costs of trading.

Thats why I like these public SEC documents so much, because they have to use GAAP so you know exactly what they mean.

1: http://www.bloomberg.com/news/articles/2015-04-21/goldman-sa...

Second: I didn't say HFT wasn't lucrative at all. I said it wasn't lucrative compared to other finance specialties. I would feel worse about this misunderstanding if the thread you were commenting on didn't include actual numbers backing this up.

Third: the ultimate reason exchanges "cater" to "HFT" is that they get paid based on order volume. That's what it means to be an exchange. Exchanges compete with each other. If nobody trades on your exchange, you don't make much money with it. Exchanges would rather you do business with them than not. It's not complicated.

Fourth: if you're in the normal demographic of HN (age 20-35 or so), you've never bought or sold a share of stock that went through a human market maker. Market-making is automated now and has been for something like two decades.

Fifth: if you're an individual buying and selling stocks for your own account by clicking buttons in an online brokerage, you are placing market orders. There aren't "minimum" and "maximum" prices. There's a current best bid, and a current best offer, and a gap between those two prices. We have decades worth of research on what automated trading is doing to that gap: it is slashing it from dollars to pennies.

Sixth: if you are placing limit orders, there is nothing an HFT or anyone else can do to "front-run" you to cheat on your order limit. Incoming orders are priced according to the price on the order resting on the book. When you place a limit order that doesn't immediately execute (in other words: when you place a limit order that is actually meaningful), that's the price your order is going to trade for, if it ever trades.

If you disagree with any of this, it would be helpful if you could lay out a hypothetical sequence of events in which this "front-running" is actually occurring. I place a LIMIT SELL order for 1000 CSCO on BATS at $27.90. An HFT decides to front-run me. What exactly happens next?

Please step me through this scenario. Exactly what do you think is happening, in what order and how?

Further, some whales (pension funds) might get hurt and that hurts your median citizen. Many hedge funds seem to have a strategy of whale-hunting. Thus dark pools and all that.

I'm not aware of a correspondingly simple way of acquiring base domain knowledge of the topics we are discussing here. Perhaps you can suggest some books?

The Complete Guide to Capital Markets for Quantitative Professionals (McGraw-Hill Library of Investment & Finance) Hardcover – 1 Dec 2006 by Alex Kuznetsov

A book on trading is prob more useful. There are many public specs of exchanges. For example:

http://www.cmegroup.com/confluence/display/EPICSANDBOX/MDP+3...

The main thing to learn is orderbook building mechanics

Note: despite the fact that retail investors virtually never trade directly on exchanges, let's stipulate that they always do.

I don't think it's a good thing that the NYSE's feeds were this crappy; that's not what I'm debating.

There is no expectation in the markets that everyone is working with the same information, quite the contrary, the markets wouldn't provide value if everyone was working with the same data. One of the chief reasons the markets are valuable is that they give people incentives to surface (in the form of market activities) pricing information they might otherwise keep secret.

> I have to think that at some level, HFT will result in higher prices paid by brokers and their clients

Quite the contrary, it has led to a race to the bottom such that I can trade for free on my phone from anywhere with internet access at spreads that are nearly nothing. Robinhood is only possible because HFT has made it so.

> and directly or indirectly by mutual funds.

Vanguard, the gold standard for low priced mutual funds, completely disagrees with you.

https://www.sec.gov/comments/s7-02-10/s70210-122.pdf

From the Vanguard comments: This roughly 25% decrease in the end value of the investment demonstrates the impact of reduced transaction costs on longterm investors.

So perhaps you didn't state a position, but you posed a thought (I have to think that at some level) or question and Vanguard rather decisively answers that question.

From a technical perspective, as orders are sent to the exchange, NYSE sends them out on a direct line to subscribers who listen to that feed. This is not private information, but it is relatively expensive, given you'd need to have servers at a colo, and pay for fast networking, boxes, software, etc.

They also then aggregate that data and send them out on a feed to the NBBO which is much more accessible to the majority of traders. Obviously, it is somewhat delayed, as they need to aggregate, and a firm that can aggregate the data faster than NYSE thereby has an advantage. This aggregated data gets set out as soon as possible, and NBBO defines a time range that they need to be accurate. For what its worth, NBBO means "National Best Bid and Offer" and they will aggregate bids and offers from all exchanges and provide the best one in a single consolidated feed.

Nanex determined that at certain times during high volume, Nasdaq's aggregate feed was delayed by up to 30 seconds, and this was cause by certain entities quote stuffing (which like I said earlier, is against the rules).

That is all that was determined. There is nothing about trading on private information (which there is none here) or front-running (which HFT firms can't do) or any of the other nonsense that gets spewed out.

If anything, other firms with good infrastructure would have been able to arb out the inefficiency and make the markets right again, quickly.

  > Once again, we detected sizable delays between time-stamps
  > in the public quote and Open Book. These delays ranged in
  > the hundreds of milliseconds. Though the delay magnitude was
  > far lower than the tens of seconds discovered during the flash
  > crash, it was still hundreds of times higher than expected.

The 30 second delay was only during the flash crash. For an aggregate feed that's not aggregating very efficiently, that would make sense, as there was an abnormally high amount of volume during the crash.

FWIW I'm getting the 'real-time NYSE data' (not the level two data but real time level one) for $2 / month on my tradingview membership. Not exactly breaking the bank

The speed and format of the data is the other, and major component of what you pay for.

You are paying for what we'd call display only quotes, meaning you get these quotes so a human can visualize where the market is. If your strategies are based on trading once a day then this is fine for you.

The other way in which you will be throttled is how many quotes you can view at once.

If you trade options you won't be happy with only being able to view, say, 20 SPY contracts at a time as there are 1000's of contracts for sale.

Ideally you'd be able to view 20 contracts per date( once a week) multiplied by 2 for both puts and calls over the next 3 months, which is 20 * 2(put and call) * 12 (3 months of 4 week contracts).

This is alot of data to stream over the internet considering that those quotes probably change 200 time a second each.

Do you work for the SEC? Because only the SEC knows what I sent them. If you do work for the SEC, then you have some explaining to do. If you don't work for the SEC, you have some explaining to do. I thought Hacker News was a place for logic and reasoning. Yet it is so easy to pick these inconsistencies out. Why do you suppose that is?

1) How much market data they give out. Typically this is broken into 3 levels:

  - Level 1, Just the top of book quote, this is the best bid and offer and their corresponding number of shares at this price level.

  - Level 2 Depth of market at each price level, same as above but you can see all liquidity offered.

  - Level 3 same as Level 2 but you can see each individual order that contributes to each price level

The other way they differentiate is the speed at which you get updates.

Have a look here to see what just the Nasdaq offers. http://www.nasdaqtrader.com/Trader.aspx?id=FeedMIPS

You can spend anywhere from $500/month to $75,000/month to receive market data from just the Nasdaq, one of 50 trading venues in the US.

As you can imagine the difference is speed. The slower methods will use the text based FIX protocol and the fastest methods give you a feed and an FPGA to parse the feed with.

It's this later feed that is getting the exchanges into trouble as these feeds notify the user of trades before the SIP can be updated which lets firms have a 300 millisecond second peek at the market before everyone else does.

Couple these faster feeds with many dark pools not reporting trades in a timely manner and new order types that allow firms to sweep liquidity at one exchange only( ie not rout-able orders) and you get a situation where latency arb can thrive.

From the article:

Specifically, we found that stock quotes from the NYSE were delayed by more than 30 seconds to the public quotation feed (chart 1), relative to Open Book, which is NYSE's expensive direct feed product used mostly by High Frequency Traders (HFT).

That indicates it was quite a bit more than 300-700 nano seconds at it's worst, by a few orders of magnitude. The chart showing the delay over time in their testing period shows a delay of 200ms routinely. That's way more than a few hundred nanoseconds.

The NYSE has always had a reputation of being awfully inept from a technology standpoint, though they have made strides in the past few years.

That particular time was made even worse by the direct feed holding up much better that their SIP feed, though the direct feed was also slower than normal.

An accurate title might be NYSE fined for having a software bug that slowed down one feed occurred during the busiest time in the markets history.

If by a special situation, you mean directly observable at multiple time periods. Here's what the article says:

To satisfy a curiosity of whether the NYSE public quote delay was unique to May 6, 2010, we ran another quote-by-quote comparison of time-stamps from the public quote and Open Book for a 30 minute trading period in General Electric stock (GE) on July 21, 2010 (see chart 2). We chose this period because there was a noticeable lag in the public quote from the NYSE versus quotes from other exchanges, allowing us to rule out the consolidation process as a source of the delay.

So they were able to find another time period exhibiting the same behavior as the flash crash at a later date, and it doesn't appear they had to look very hard (but there aren't a lot of details on that). Thirty seconds is a crazy amount of lag, but that was during the "special situation". That they look to have routinely had lag well into the hundreds of milliseconds is more troubling, because there's no longer a special situation to point towards as an excuse.

To be fair to you I originally posted nano second when I meant millisecond. I've corrected my post.

As I've said before, I'm not at an HFT firm but we routinly see latency changes in messagesfrom all excahnges and dark pools.

Just like Google can't guarantee each query will be served in 250 milliseconds, each exchange can't guarantee every message will be delivered in nano seconds.

I'm good with this.

Ah, that makes a big difference. :)

> As I've said before, I'm not at an HFT firm but we routinly see latency changes in messagesfrom all excahnges and dark pools.

From separate feeds to the same exchange? My understanding from the article is that that is not allowed:

A crucial sub-ruling in the regulations prohibits exchanges from giving stock quotes to special groups faster than to the public.

So as long as you're measuring from the start of the request, not the end, you shouldn't see delays in the hundreds of milliseconds to the same exchange, regardless of feed. At least that's how I understand it, but you sound like you have more practical experience, and may be able to correct where I'm misinterpreting something. Is that not how it works in real life?

This is not about a time when some technology broke under extreme load, this is about a broken system that persisted in that state for an extended period of time, and it was broken in a way that made it no longer compliant with the law.

Err, No. You can parse FIX with an FPGA if you feel like it. It is up to you how you consume the data. The differential will be CoLocation vs dedicated line vs internet delivery on a base level and then differing subsets of feeds to varying liquidity pools.

So really you are paying whomever owns the DC for the privilege of a seat in the same DC as the exchange and for the cross connect right into their cage.

That's true but no one said you couldn't. What I said was that FIX based feeds would be slower than binary ones. I stand behind this.

I also said that some binary feeds come with an FPGA card for parsing them, you are of course free to use your own FPGA to parse a fix feed or binary feed, but the binary feed will be faster due to it being smaller.

>The differential will be CoLocation vs dedicated line vs internet delivery on a base level and then differing subsets of feeds to varying liquidity pools.

This is wrong, but its not your fault, its clear you aren't in the industry or really understand it at all and that's ok.

Colocation is part of what you pay for, but the latency and bandwidth of the feed itself also is part of what you pay for. You'll also pay for the type of data you receive and the SLA contract that dictates how fast you'll receive your data at the 99.99% level.

Feel free to mail me if you'd like to learn more!

You say that like it's a bad thing. My understanding is that an "intermarket sweep order"[1] (aka ISO) is essential to the proper functioning of the market. Otherwise, here's what used to happen all the time:

- the "true" market for XYZ is $10.00 bid, $10.01 offer. (Let's ignore the mantra of relativity that "simultaneity does not exist").

- in this case lets assume NASDAQ prices currently reflect that "true" market, and prices disseminated by NYSE are stale.

- a buy order arrives at NASDAQ, willing to buy at $10.01

- NYSE is supplying stale data, claims it has stock for sale at $10.00

- without ISO, NASDAQ can't allow the transaction to occur at $10.01, but is obligated to respect the stale data it is receiving from NYSE. NASDAQ must reject the order, or route it to NYSE.

- SUCKERS! the NYSE quote was stale, but the order was not transacted on NASDAQ. And, there is no longer any stock for sale at $10.00 at NYSE, so it won't be filled there either!

Here's an article that claims that ISOs are a form of protection against high frequency traders:[2]

   Brokers that don’t use ISOs in fast market
   conditions disadvantage their clients and
   subsidize the profits of HFTs

[1] https://en.wikipedia.org/wiki/Intermarket_sweep_order [2] http://tabbforum.com/opinions/why-hfts-have-an-advantage-par...

> You say that like it's a bad thing.

Umm I said no such thing. Please don't make thing up.

The order types are useful to market makers and help prevent locked markets, which REG NMS prohibits.

I use them every day and I"m not at an HFT firm.

IMHO markets should do something similar: create something like 30-second barriers for trades. Trades that cross the bid/ask are carried out. Trades submitted cannot be canceled until the next 30s cycle. Trades outside of the range (bid too low/ask too high) stay open. The open trade book is sent to everybody every 30s.

Allowing µs-range trades (and particularly adding and removing bids and asks so fast) is asking for trouble, it is obvious that at some point of time two or three HFT robots will react on each other's trades and do something veeeeery stupid.

Meanwhile, if you're the kind of trader that is impacted by trade latency (read: a competitive market maker or decently well capitalized prop trading firm), there are already venues you can trade in that are quantized. The problem is that nobody wants to trade in them.

edit: "Across Wall Street, less trading is taking place at 3 p.m. or, in fact, most any time but the opening minutes and the final half hour." http://www.wsj.com/articles/traders-pile-in-at-the-close-143...

Think about the billions (trillions?) of trades that occur on minescule timescales now and how often we have flash crashes. Sure, HFT algos interact with one another constantly, but flash crash investigations have uncovered both human (manual entry) and algo error as their root causes. Part of what makes flash crashes more severe is when HFT safety protocols kick in and they remove their orders from the market, causing liquidity to dry up and exacerbate the crash.

"If the average or base quote rate is around 10,000/second, then it only takes an additional 10,000 quotes/second to reach the magic 20,000 quotes/seconds where a corresponding delay is seen in NYSE quote from CQS."

This seems like a bug related to system load, so a $5M seems pretty harsh unless they knowingly ignored the issue for the benefit of their high-paying CQS customers.

- Is HFT a healthy or unhealthy part of the economy? Meaning does it help flatten out the highs/lows, or does it emphasise them?

- Aside from a micro-tax on trading (e.g. 1c/trade) is there any other mitigations/solutions to HFT?

But volatility isn't the only axis on which to measure the health of the markets. The advent of electronic automated market making is also tied to drastic reduction of spreads, which are a tax that anyone trading has to pay to the market for the privilege of buying at the market.

A lot of people talk about spreads as if they were a meaningless abstraction, in sort of the same way as they talk about "liquidity". But in the 1990s, crooked human market makers (read: all market makers) rigged spreads to skim extra money off trades; look up the "odd eighths scandal". That doesn't happen anymore, and, more importantly, spreads (and their attendant cost to trading) are razor thin.

That's not the only thing automated electronic trading does to improve the markets. For instance, in the 80s and 90s, simply placing an order with a broker was extremely expensive. Today, there are no-fee brokers. Why? Because automated trading systems purchase the rights to internalize orders from retail brokers; if you're an automated market maker, retail order flow is practically risk free revenue, and all you have to do to earn it is hook your trading system up to brokers.

1: https://news.ycombinator.com/item?id=3852341

But it's often not "real" liquidity. Let's say there's 100 shares available at $100.00, 200 at $100.50, etc. Such that there's a total of 5k shares on the book at this time at any price. But we want to buy 10k shares.

The average fill price assuming no intervention and only marketable orders would be (100$100 + 200$100.50 + ...)/5000.

An intervening HFT firm will only be adding useful liquidity if the average fill price drops somehow. Shittier algos will lose some money by lowering the average fill price so they match your order and get the liquidity rebate from the exchange. This adds true liquidity. But intelligent market makers generally will just add identical liquidity to what's already available but at 1 or 2 cents higher (and a few slots higher on the order book). As a result, you just end up paying more for the same quantity of shares. That's not true liquidity because it's really only triggered by your order, doesn't move the price significantly, and is limited to exactly the size of your order.

Why wouldn't the sudden appearance of a big buyer that wanted more liquidity at near levels than is available indicate the price should go up? That's supply and demand. Am I missing something about your argument?

I'd rather we found a way to get everyone to work on drug discovery and civic engagement, but unfortunately, we don't generally put regulatory brakes on technologies simply because they attract developers that would otherwise be working on our pet issues.

You understand that exchanges compete with each other, right? The markets are now decentralized; there are like 15 different lit exchanges on which you can trade stocks.

"Most HFTs run a market making strategy. What this means is they play both sides of the table - they take no position on whether a stock will go up or down. Instead, they try to offer securities both to buy and sell. If you want to buy, they will sell to you at $20.10. If you want to sell, they'll buy from you at $20. As long as their buys and sells match don't get too out of whack, the HFT will collect $0.10 = $20.10 - 20.00."

That does provide more immediate trades for both the buyer and seller, which the pro-HFT camp calls liquidity. Unfortunately it's usually not needed - both the buyer and seller were already in the market and willing to trade, and now they both just had an additional $0.10/share sucked out of their trade by some HFT.

The problem is, how much immediacy do traders actually want? HFT provides immediacy in microseconds; its needed for continuous-time auctions because buyers sellers usually aren't in the market at the exact same microsecond interval (while the HFT is there at every interval).

Frequent Batch Auctions[1], where orders are matched in batches with single-price clearing (call auction style) - just like the opening and closing sessions of NYSE, but much more frequently with shorter periods of say 30 seconds, 1 second, or even 100ms per batch. Their models show that frequent batch auctions will reduce spreads even further (because it eliminates mechanical/latency arbitrage rents). In theory, retail traders should get even better prices (because the spreads will be smaller) if they are willing to give up microsecond immediacy for 1 second order matching.

http://faculty.chicagobooth.edu/eric.budish/research/HFT-Fre...

I don't particularly have an opinion about the Budish paper, other than it has some serious impacts to globally traded products and it is way more disruptive than people seem to suggest.

I do worry that fundamentally altering the way the markets work, because some participants don't understand order types is particularly troublesome.

The complicated types (hide-n-slide, NBBO pegs, and conditional pegs[1]) may help some players at the expense of others. The only players who wouldn't benefit from batch auctions (according to Budish et. al.) are latency arbitrageurs.

1. https://mechanicalmarkets.wordpress.com/2015/10/05/iex-peg-o...

Thats not quite true. Price improvement happens all the time on limit orders, but I'll concede you'll pay close to it if it fills.

> The complicated types (hide-n-slide, NBBO pegs, and conditional pegs[1]) may help some players at the expense of others.

This is true of limit/market/stops as well.

The Budish paper is very interesting.

What we don't know is what it would do to the spread, which is very important especially given that it will be harder to predict the clearing price. It is also a major change to the markets, virtually impossible to implement and its not clear to me what that risk is buying us, especially given that I fully expect most participants to go through an intermediate first.

A) a counter-party never showed up B) a counter-party only showed up when the stock dropped so you end up buying on the way down.

Taking this risk is a perfectly fine thing to do and many people do this. Others would prefer to pay the market maker a small fee for providing liquidity. That fee is not being sucked out by some HFT. It's a fee for providing a service.

And BTW, thanks to automation and sophisticated trading algorithms that service is provided at EXTREMELY low cost these days. Bid/ask spreads are tiny.

It's worth noting that after the Canadian regulatory authorities added a per-message fee, bid-ask spreads rose, liquidity dropped, and it was retail traders who suffered the most. http://marginalrevolution.com/marginalrevolution/2014/04/a-s...

(Somewhat akin to using game engine bugs for artificial lag creation in MMOs on the opponent side).

I asked my accountant friend about whether he'd ever consider whistleblowing if he saw wrongdoing. His decision criteria was based on whether he'd A) Ever be able to get hired again or B) If the size of the prize was so big he'd never have to worry about it.

[1] https://www.sec.gov/News/PressRelease/Detail/PressRelease/13...

[2] https://en.m.wikipedia.org/wiki/SEC_Office_of_the_Whistleblo...

When the SEC called Hunsader in June 2015 to tell him about the award, he said to them "I would have accepted $1 if you simply acknowledged me at the time."

Was he initially ignored?

A NYSE Speed Bump You Weren't Aware Of https://www.iextrading.com/about/press/op-ed/

Thanks for correcting me.

NYSE and NYSE ARCA are different exchanges.

So what are your options? Do you not contribute to a 401k or IRA, pay the full tax rate on all of your income today, and invest in real estate or some other investment?

There's a couple of inherent challenges.

The SIP that aggregates the data is by design a separate system from the trading systems. This is for obvious reasons, as every participant provides their quote to the SIP. So no matter how fast the SIP the is, even with a 99.999% latency of 20 microseconds, the SIP will always be 20 micros + travel time to the SIP (1 to 500 to 7000 micros) as opposed to a direct feed. Nasdaq is in Carteret. NYSE is in Mahwah. BATS is at NY4 in Secaucus, etc. So direct feeds will always be the fastest, even with a SIP that can defy the speed of light a little bit.

Nasdaq does actually read marketdata directly from other exchange participants and falls back to using the "slower SIP" data as a fall back in the event of any issues. This allows for the latest quotes to be respected.

The challenge with the SIPs being hundreds of millis behind is partly due to the criticality & member-driven consensus system in place. Each of the marketcenters NYSE, Nasdaq, Bats, etc have a stake and voting power on how the SIP would be run & operated. Since it's a system that "works" there's really no motivation to continually improve the technology/latency of the SIP. There's also a fear to change the SIP and lead to it going down. If a SIP goes down, by rule, a regulatory halt must be issued and all trading in the securities that particular SIP handles must stop. This is what occurred during the 3 hour outage several years back at Nasdaq. The trading system was fine, but since the SIP was down, everything had to be halted. Halts are detrimental to business since the respective marketcenters make money on trades. Halts = No Trades = Lost Revenue, so there is a lot of motivation to not halt by all the member exchanges.

That ambivalence to a tech refresh has since changed since the scathing critique of Flash Boys and the Nasdaq SIP is in process of being refreshed to the same architecture as the trading systems, which should bring median latencies of the SIP down into the tens of microseconds and 99% latencies into the hundreds of microseconds. Current medians and 90th percentiles are at 450/810 microseconds, which is considered an eternity.

http://www.wallstreetandtech.com/infrastructure/nasdaq-omx-w...

http://www.utpplan.com/DOC/UTP%202015-Q4%20Stats%20with%20Pr...

On the technical side, one of the challenges with synchronizing packet delivery between a private & public feed is that the absolute fastest NICs available today with kernel bypass can get a packet from wire to memory or memory to wire in about 2-3u. So every hop of a computer costs a minimum of 4-6u. Usually this is Order Port Ingress -> Matching Engine -> Order Port Egress. An FPGA NIC that does the IP Header processing and delivers data to memory can do wire to memory in about 1-2u. One-way PCI-E latency is around 800 nanos. So there's about a 500-1000 nano budget to process a packet in the FPGA, then another 800 nanos to deliver it over PCI-E to memory to be processed. The fastest cut-through switches have a port-to-port latency of around 250 nanos, so switch hops are pretty negligible in the overall latency of a system, but they do play a part. Anyway, making sure that everything is super duper optimized and running as it should be in a semi-continuously deployed environment makes for interesting and quite crazy/bull-shit level challenges.

My question is about the business side of Nanex. Are sales of your products such as NxCore enough to sustain the company? Do you also actively trade? And what opportunities do you see currently around data science and machine learning for entrepreneurs in the quantitative finance world?

For a better version of "Flash Boys", read "Dark Pools"; Dark Pools also takes a gimlet look at high-speed trading, but also does a great job of documenting the birth of the NASDAQ ECNs and the transition from the early-90s centralized markets to modern, decentralized, electronic trading.

If you want to get into the technical detail, the best book is still Larry Harris' _Trading and Exchanges_; it is almost exactly the "TCP/IP Illustrated" of trading.

> a high quote rate from any NYSE stock will cause a corresponding delay in all NYSE public quotes! Anyone with this knowledge can easily cause latency on demand across many stocks, by simply blasting quotes in any one stock

Or another link from this thread.

> U.S. stock exchanges have invested huge sums of money creating two-tier markets - building and offering faster data and technology infrastructures at a price that only a small niche of traders can benefit from or afford, while at the same time continuing to offer slower products to everybody else.

https://www.iextrading.com/about/press/op-ed/

So: not delaying public quotes is table stakes. What could Google do better than that?

I'm not so much challenging you as I think it's an interesting question, and would like to see it hashed out in technical detail.