Late deals canaries: Cheap Last Minute Holidays To Canary Islands 2022 / 2023

My Orc Trader hack wasn’t a perfect solution, but it made a few people happy with trades that weren’t possible before. They even looked happy when they saw these deals in the reports with their own eyes. However, overall it was only a minor success. In the future, it turned into hundreds of manual rules for finding deals, recouping only the costs of the operator with its trading. But at least I had fun.

Focusing on “

< 2ms”

Let’s remember the “< 2ms” condition written on the window pane. I wrote an article suggesting a faster and more sophisticated way to achieve even lower latency. A team of three programmers, including me, had to implement it. The first choice fell on the Korean stock exchange. There was no need for price caching in this project, since the European option pricing model is not much more complicated than simple interpolation, even in multiple dimensions.

At this time, processors with a frequency of 2-3 GHz were just beginning to appear. Our performance metric was a little off – we mostly measured latencies inside our code rather than those inside the operating system (which are actually much longer.) This still worked when the data came in at millisecond intervals, because at that time the network stack spent on packet processing from 50 to 100 microseconds. The main thing to remember here is when the processor can execute more than billion operations per second, then over a million instructions fit into these two milliseconds. To be honest, this is very, very bad code if this million instructions is spent on simple price calculations. If it takes you a million steps to decide on a deal, quit immediately. Here every nano second is important.

It turned out that I was assigned to a team that was not very cohesive, and as a new employee, I had to be careful. There were only three of us in the development department. One of the team members (who was actually an excellent programmer) apparently got tired and began to spend all his working time on games and forums about video accelerators. “Mr. GPU” had previously written the company’s key libraries for multithreading, and they were not bad, although they suffered from design problems. All synchronization was done through mutexes and any operation required locks. I ended up using the C++ Boost library again. “Mr. GPU” apparently did not enjoy the project and looked like he was on strike of some kind. He was no doubt annoyed at having to work on my idea. The other guy was less capable, but still minimally competent. He usually ate a lot and constantly went to the meeting room to watch the next movie on DVD there.

But we had to somehow manage to overcome these two milliseconds. As it turned out, I needed to write more code than expected – I had to stay up late several times. “Mr. GPU” did the job of integration, while “Remote Dude” did nothing at all. As a result, our new system showed delays of just under five microseconds. I thought it was a huge success. However, this figure did not take into account network delays. Definitely, they are now much more important than a month ago.

One of the other employees congratulated me. From him, I learned that the bulk of IT people (including him) considered this an impossible task. Apparently, they hoped that I would fail. The changes I brought were not widely accepted, but I received a small bonus without a significant increase in salary.

I used to expect to stay in this job for twenty years, especially after the dot-com bubble. But now it is clearly time for me to move on.

Looking for a new job

I contacted several HFT companies and shared my thoughts on what could be done. One good Dutch trading company listened to me very carefully, but offered nothing. Years later, I learned that their team started to implement something like what we talked about in the interview, the very next week after it. I heard that the approach was successful. They called it the “implied base method”, in which everything came down to predicting the theoretical price of a futures or underlying asset. I like their name for my method even though it looks like a cargo cult. Well, that’s the way financial companies usually do business. Despite their deceit, I was flattered.

I then spoke to an Australian company with a team of about a dozen developers focused on trading on the ASX exchange. Their platform was basically even slower than Orc Trader, taking many milliseconds to calculate. I asked for a meeting with the head of the company, who was much more successful and much younger than me. He was one of the super-high paid guys whose income had to be declared under Australian law.

Their business didn’t look “totally clean” and he warned me that the money they earned abroad was not always legally easy to return to the parent company. Negotiations with this firm were a waste of time – I don’t want to be a black hat. Gray “Finding gaps in the rules is a fun game, as long as it’s ethical. However, these are the red lines we shouldn’t cross. There’s no fun in winning by cheating, but more importantly, your reputation matters. ” People always underestimate the cost of scandals and I didn’t want to risk it.0003

I also liked one of the investment banks that I spoke to. It was a branch of a large Wall Street firm. I have already worked for several years in this kind of environment. The entry threshold seemed pretty low—they didn’t want to make money, they just wanted to generate turnover at no cost, so their finance department could serenade the affiliate. This seemed like a good option to me and I was sure they wouldn’t mind generating turnover as well as making extra profit. Profits rarely cause complaints.

The good thing about my previous firm was that they could lay off even earning employees. Someone broke the rules by making a few transactions that were not authorized. They made the company money, and I was impressed that they got fired anyway for violating the rules. This correct management position is extremely rare, especially in finance.

Then, quite unexpectedly, another possibility arose.

I once had lunch with a guy I knew in a computer science course at the University of Tasmania who was now working for ITG. He suggested that I speak to ITG as they were private traders. This was news to me – I thought they were just brokers. I met with their CEO, Asia Pacific, and he spoke in detail about their arbitrage platform, which works with USD and CAD denominated stocks. Then I met US CEO Ray Killian while he was on a trip to Australia. The meeting went well and there seemed to be no more barriers to the development of high frequency trading at ITG. The Asia Pacific CEO plus their chairman, who was also a board member of NYSE-listed ITG Inc., ended up taking the impromptu business plan and risk calculations to the US for a board meeting.

What happened next I don’t know – I was told three completely different stories over the next few years. According to one version, the board approved the business plan, but Killian played everything back. According to another version, some members of the board did not approve the project. According to the third version, the business plan did not reach the board members at all. I don’t know which versions were wrong (perhaps all at once), but it didn’t matter. The guys from ITG decided to do without their employer and invest their own funds. It wasn’t quite the job I was looking for, but if the business went well, I could earn about a quarter of the shares in the new company through options vested over several years. It felt like a victory.

Birth of an HFT company

With an investment of AU$1.5 million, it’s time to start working alone with a laptop in the office again. I set up a few servers and workstations, installed 64-bit SuSE Linux, hired three programmers, and the clock for the investor-paid period started ticking. It turns out that I’m not a very good Linux system administrator. Soon I got tired of doing poorly admin tasks and entrusted them to a freelancer. Now there were four and a half of us. But it turned out that 64-bit SuSE Linux, especially on AMD processors with their Hypertransport bus, was not ready for prime-time workloads, so we returned to a 32-bit build of RedHat/CentOS. It was a wise decision by “Mr. SysAdmin” that moved us forward a bit.

Four months later, we took a look at the prospect. We had to choose whether to launch on our ASX or go to Korea. Three Korean exchanges, KSE, KOSDAQ and KOFFEX, were just merging into a single Korean exchange, KRX. We chose Korea, although both the business and technical environments there were more complex and risky. However, there were more features there too, plus simpler math in pricing models and a small number of tools promised a more lightweight codebase.

A funny thing happened at the same time. When creating a new system, you usually try to come up with some naming conventions to make your life easier. I decided to use ISO codes or part of them for standardization. However, due to the merger of Korean exchanges, no ISO code has yet been assigned for the new exchange. I contacted the responsible ISO team, who turned out to be the only excel spreadsheet guy sitting anywhere in Europe. He decided that the new code for the combined exchange would be XKOX. I gasped and wrote to him that this combination does not sound very euphonious in English. Perhaps a better idea would be to use the obvious codes KRX or XKRX? We laughed and so XKRX became the new ISO code. So now I can claim that I’m the one who created the ISO code for the unified Korean exchange.

In those early days in Korea, traders, even high-frequency ones, could not get direct access to the exchange. You had to go through the broker’s gateway and use their specific API to connect. There was no collocation at all. Among clients, there was a large difference in delays that was not controlled by the broker, simply due to different locations and interaction technologies. Futures and options were traded on the KOFFEX exchange. KOFFEX itself was located in Busan, a port city a little over three hundred kilometers from Seoul. Derivative contracts for the KOSPI 200 index had a small denomination, but a large lot size. Because of this, they were the most volatile derivatives in the world, with a tenfold lead over the rest. The EUR/USD contract on the CME was number two and the options on the KOSPI 200 were number one. This turned out to be too much for KOFFEX, so the main trades in these instruments were carried out on their behalf on the KSE exchange in Seoul. So everything we needed from the exchange was located at one point in Seoul, although it itself was based in Busan.

By this point, I already knew that the networking stack introduced the most significant latency that I could manage in our new implementation. I found a hardware ethernet protocol converter in InfiniBand that introduced a latency of 3 microseconds. It was made in the form of a module that was connected to the TopSpin InfiniBand switch. AMD’s Hypertransport (HTX) bus, an alternative to PCI Express, was more modern, resulting in lower internal latency. I bought several Pathscale HTX Infinipath NICs and HTX capable motherboards (they were only in alpha versions at the time, for example one of them had serial number 0x0000045!), put all these motley parts into a large case and got an expensive, but an extremely fast networking stack that I was very pleased with. He looked, of course, somewhat awkward and clumsy, but he was one of the most competitive in the world. For a few more years, regular NICs couldn’t match my configuration.

Choosing a Broker

After that, we went to Seoul to find a broker and a place in the data center. I went with one of the investors, a man named Bubble, who then became our employee. At first he was torn between ITG and HFT, but later he became completely at our disposal, since both the CEO and the chairman of the Asia-Pacific region, as well as other ITG employees, were our investors. It was he who was instructed to book a hotel in Seoul. It turned out that he had chosen a motel with hourly rates, seduced by a very picturesque artistic description that obscured the true picture – rundown, dirty, rooms with peeling air conditioning, a virus-infected computer. Not a very good start. Several brokers stopped by and expressed surprise at our area. It’s awkward, but funny. The trip turned out to be quite productive, we made a stunningly incredible deal.

Good deal

Muppetz Securities offered us a DMA service. The big firm I left didn’t have direct access. Before that, I was looking at an HP/Compaq/Tandem system at an old job, as it was one of the ways to get into a particular broker’s internal network to get closer to the exchange. It cost about US$500,000 for a small box created through HP’s custom in-house application development program. Seductive, but not for a business that is living its last days. The DMA service offered to us now was implemented on the basis of the X.25 protocol, which was then innovative for me. This broker deal was quite expensive, but since real direct access was unheard of before, I signed the contract a few days later.

At that time, I met “Mr. Ree” from the local company RTS, who was passing through Seoul doing his own business. We helped each other in finding brokers in Korea. We used the RTS internal information system as a back-up store for our transactions so that shareholders can compare our data with those broadcast by independent vendors. You might think that the Orc platform would be the logical choice. I was familiar with Orc and at this stage it had the most user friendly interface. However, RTS had a much better API and cost an additional 10% of the monthly cost.

Back in Sydney, I was looking at all this expensive InfiniBand equipment. Now it was useless – we were going to switch to X.25. An unexpected but pleasant turn of events. Then I got a message from the broker Muppetz Securities that he was canceling our trade despite the contract already signed.

Some time later I learned that the Muppetz had gone to the big office where I used to work. X.25 was a very limited and rare resource. Why give it to some small startup HFT firm that isn’t even trading yet when you can go up a level and make a deal with real traders? Smart move, albeit unethical.

It turns out that saying “no” has weakened the brains of some brokers. I believe a lot of HFT firms have been thinking about how to get direct access. It must be time. However, the cancellation of the Muppetz contract led to other, much better offers. One from Mr. M, a real gentleman and just an amazing guy with a good team. Its price was much better than the Muppetz and although technologically this connection was a brokerage access, the location was amazing – the connection point was right in the exchange building!

I had no idea you could trade there. It turns out that this firm had a dusty, plain, air-conditioned room in an annex to the old KRX building. This old building, in contrast to the new building recently erected on campus, was connected by a corridor to the main tower, where the exchange engine was located several floors underground. This room was used as a place for emergency terminals. There was a small table in a narrow room with terminals with old CRT monitors. If the whole world were to end, employees would have to be sent there to enter the keys. This room was not adapted for a server infrastructure. Would we be interested in this little crappy room, which is the best they have to offer at the moment? I tried to hide my excitement by saying yes.

Muppetz’s X.25 solution was technologically faster, but the connection point was in a building outside the KRX campus, which added latency. Someone was going to pay a big premium for a solution that wasn’t as cool as the one we just came up with.

X.25 and technology

The X.25 exchange infrastructure platform was interesting. You could count on a T1 or E1 class line (the exchange had both options and used them alternately) from Koscom, KRX’s main technology and network provider. Within the line, virtual channels were allocated at 64 kilobits per second (kbps) according to the principle of time alternation. Interestingly, because of this interleaving, you essentially have parallel access to the entire channel at the same time. The broker has a limited number of lines, but many parallel (though slow) channels.

I did some testing and first bought X.25 NICs from a UK vendor. Popular enough to be worth mentioning, they seemed to be of industrial quality. Years ago, I was developing an information system for Citibank’s car leasing accounting system, and it used the X.25 protocol. The entertainment that this development gave me was only possible because I had an unfortunate youth in the pre-internet days – BBS, FIDO, Z-modem, Y-modem, that’s it. Most of the time at Citibank was spent fiddling with modems and other equipment just to get it all to work together. After that, everything was easy. But not with these network cards. Neither we nor the card vendor could get them to work properly. In the end, we had a conversation in raised tones, after which the seller simply stopped picking up the phone. We then bought some cheap NICs from the Canadian company Sangoma, which worked right away.

We set up our server rack in our dusty workroom. I came with hand-built servers in a bag. Communication lines have not yet been allocated for us. The channel to the exchange will appear a little later, but now we urgently needed market data to prepare and set up our infrastructure. HFT requires a lot, a lot of data. The exchange solved this problem by running the twisted-pair cable through the air outside the building and leading it through a hinged window into our room, which we had to keep open at all times. When it rained, a little less water got into the room than one would expect. Twisted pair is connected to T1 modems. Modems, in turn, are connected to serial ports. Giant X.25 connectors, a legacy from the sixties, protrude from our Sangoma boards. Slow beats are big beats.

Soon the line from the exchange was connected to send orders with even more gigantic X.25 reciprocal connectors. We had modules for working with Postgresql (soon replaced by versions with flat files), a trading engine in C ++, a service for calculating volatility, etc. Instead of leased lines, as we initially agreed, we used regular Internet channels. We took three providers who promised geographic independence and assigned routes to provide redundancy. Much cheaper, higher throughput and more reliable than leased lines. It was unusual at the time, but many do the same thing now.

(Fal)start trading

Nine months after launch, we finished testing. No certification from the exchange was then required. At that time it really was the land of cowboys. We started trading – there were 85% executions on options! This means either you can’t price options correctly (and anyone can sell you contracts at bad prices), or you’re always first in line and you’re doing well. We had mostly the latter.

But we couldn’t make any money.

Our original plan was to simply buy juicy options and hedge them with futures. We spent almost a million dollars developing a platform for this, and the balance of half a million remained in the brokerage account as collateral for our trading, we could not touch it. We may have time to grab the option, but all profits will evaporate if we cannot hedge it properly. We were constantly working on it, and things improved little by little, but the problem was not completely solved. I spent the night in the office and the rest of the team also worked hard. The performance of our trading engine was insufficient. In the end, I spit on everything else and, straining (as they say, “catching a bullet with my teeth”), rewrote it in 72 hours of continuous work. After that it got much better, but we found a new bug in the new trading engine. Internal application delays masked it, but it was noticeable in external measurements. The thread that wrote the logs was consuming the main thread time quantums, which cost us a lot of time. The correct solution was not easy, since we would have to set and check global flags. I just decided to give the main thread a real-time priority with the corresponding deprioritization of the others, and… hooray! – problem fixed. Although this was essentially a crutch, we got a stable delay, measured in two-digit numbers in microseconds.

Deriving an implied volatility (IV) curve has always been an art. In my old office, traders used a manual IV fitting process, and every five or ten minutes (or hour) they adjusted their curve to match the market. A smart quantitative trader would eventually build a neat and reliable automated process. Here, too, I wanted the IV correction to be fully automated, but took a different approach. Using quadratic curves for valuation didn’t work for me, as it often resulted in an uncomfortable gap at the cash out point. We used more appropriate splines from the R libraries, added some small median filters, and some other statistical tricks. Now we calculated and published the IV curve every second. Later we even switched to quarter- or half-second intervals, but seconds were enough. This somewhat improved the situation and our option valuation seemed to be very close to reality. However, as a rule, we did not make money (or made very little.) It was time and running costs that killed us, not trading losses.

We had another remarkable puncture. Mr. L has made some changes to the futures hedging mechanism that we have rolled out. The second we turned on the system, it instantly bought and sold on the market about a thousand lots of futures in a few seconds. We should have killed him immediately. The company could have died right then, but we were lucky to jump out of position.

We tried, but still couldn’t make money. And these were no longer delays. The prices weren’t terrible. Our futures hedging mechanism was simply not efficient enough. We didn’t have the capital to “draw” the levels and use them to improve our financial performance. We tried to do passive market making on futures before trading options, but it didn’t matter. Nine months passed, this child’s gestation was complete, and sleeping in the office was less and less beneficial. It’s time to pray.

Zero Strategy

When you trade on the stock exchange, you are considered lucky if you get a part of the spread in your favor. At that time in Korea, ATM and lower options had a tick size of 0.01 pip. These 0.01 were worth about a dollar per contract. When trading a high-frequency robot with a series of options at once, if you are lucky, then you exit the transaction with an average of ten or twenty cents. One of my goals in this job was to surpass the turnover rates that I had seen before. I remember that the record in the last place was 1.13 million contracts per day. If you could trade a million options and take 10 or 20 cents each, you would be very happy. In theory, there is no difference between practice and theory. It was quite difficult for me to realize this.

Then I remembered a conversation between one of the traders, Mr. B., and someone in the kitchen at the old firm (he ended up running the firm, by the way.) Mr. B. was asked if he bought on a good trading signal and made a profit , why doesn’t he immediately close the position? He replied that if everything was so simple, he would earn much more money than now. We had excellent latency with a good location and a direct X.25 link. The technology worked, the valuation model gave correct prices. Maybe we could do the impossible?

I spoke to the shareholders and told them that I needed another 25,000 AUD to get through the month and try something new. The main investor issued a loan in the amount of 25,000 AUD. Mr. F., one of the original employees, coded what we called the Zero strategy. Everything was simple – not to have a directional position, to wait for an option to appear in the order book, whose price differs from our estimate by at least a tick. Then hit him. Wait. Then discard it and take the tick profit. Could it work?

We will live!

Our monthly trading income has become a healthy six digits. The Zero strategy worked. We increased the trading capital so that we could buy 500 or 1000 ATM options and then turn around and dump them back. There is no smart market making here. We were just being an aggressive taker.

Being a market taker can be more problematic than being a maker. As a market maker, you have to get out of the way of the market when it goes against you, but you have to be stoic enough not to get too flighty and lose a valuable place in line. As a rule, your price threshold, after which you need to run out of the order book to prevent a loss, is less than that of an aggressor who uses a trading advantage. It looks like a strange game in which the aggressor actually has a delay handicap, since he is only ready to move after the maker has already made his move, all other things being equal. Or is it like a 100m sprint where the maker has a 50m lead and you have to chase him somehow. This analogy pretty much describes the Zero strategy. A direct connection via X.25 with wires running through a window and location on campus was enough to overcome the naivety of our strategy in those early days. Fortune smiled upon us.

It was a kind of entertainment to watch the money appear in the bank account by itself after making a profit. It was a little surreal. We trade index options in Korea and Australian dollars appear in a bank account in Sydney.

Some of the money I made went into a home cluster in our newly rented 200 square meter office on Hunter Street, lots of small mini-ITX computers on a cheap hardware store shelf, RAM boards, ATX power supplies, and a network switch. Implied volatility modeling, backtesting, and the strategy itself have all improved with a few hundred additional cores.

The next step in our trading was the transition from the Zero strategy to a full-fledged hedging trading strategy. Instead of trying to hedge the position through futures, we turned to hedging with other options. We did not want to carry positions through the night, but were happy to hold them during the day. The Zero strategy kept a cumulative exposure of a few seconds throughout the day. The new strategy forced us to hold the position all day. This idea worked very well. About two-thirds of the KRX market consisted of options with low deltas and fairly static pricing. It was not unusual to see over a million lots of bids or asks in tiny options. However, buy these cheap options and it will be very difficult for you to dump them before the market closes. So we just focused on options with tighter spreads and larger deltas, tick size 0.01 with prices below 3.00. They accounted for about a third of the average market volume. We ended up growing to a typical 6-7% market share and sometimes over 10%. It is amazing that we traded about a third of the volume of the market we focused on. A normal piece of cake. Overcoming the previous speed record took place.

Dealing with wires

The idea for the first significant hack came from looking at a twisted pair cable that was run through a window. The modems for the T1 and E1 lines transmitted this signal through clumsy connections leading to an even larger box, which was then connected via huge connectors to the serial cards. A little research and an obvious plan was born. I bought some new Sangoma cards with T1/E1 channels from NetOptics connected with twisted pair and installed this assembly in Korea. When I was asked what it was for, I answered that it was for monitoring the state of the line and measuring delays. This statement was technically correct, but incomplete.

The new Sangoma cards supported mixed mode, where the cards simply accept all the bits from the channel. I began to analyze them and understood what the flows looked like. X.25 virtual channels have been grouped into market data channels. They tended to follow each other in succession. There were other channel groups for puts, calls, and futures. I had to parse the HDLC layer myself, but soon I had decoded packets corresponding to the KRX channels. This bypass of the modem and network device saved about a millisecond.

This was not unfamiliar territory. The Eurex exchange once provided a closed API – and it was hacked and the exchange protocol was reverse-engineered. After that, the official document on the technical regulation of membership on the exchange was adjusted so that it became a violation of the rules. Then this practice was stopped, but in Korea there was no such rule. Similarly, when ASX opened its first small co-location facility in Bondi, I noted that several client firms had 1U Dell servers installed, with no modems. Reverse engineering of the Nasdaq protocol wasn’t as uncommon as I had hoped – there are a lot of smart hackers out there.

Cut off and discarded

Things got even better in Korea. Now that I had bit by bit streaming (or at least byte by byte) I could see that the packets were wasting a lot of time on the wires. milliseconds. A common technique in network monitoring is slice, where you simply cut off the first part of a packet. I implemented this with my HDLC layer traversal code and now we were way ahead as we needed the bid and ask fields at the beginning of the quote packet. Save 100 bytes at 128Kbps and you’re ahead by 6.1ms. As the specification parameters grew and the KRX channels accelerated, our advantage decreased, but still remained acceptable. We were now submitting our orders before the input price packet was fully accepted.

Sometimes the exchange swapped virtual circuits, and we were left with a slow cable until I decrypted new numbers. The rest of the infrastructure also slowly improved. Now we had cables running not through the window, but through cable channels from the building’s main switch. We also hung a small air conditioner over the window, which was very necessary, because the temperature in the rack was constantly rising. Our workroom was not a real data center. Then our bank sold its derivatives business and after a while the new owner decided that we were too small customers and ordered us to leave our office. Later I found out that he had his AI sales group in Tokyo, who really wanted our room for themselves.

Another try

We struggled to find a new home. We focused on finding several points at once so that the business would be a little more sustainable. The X.25 protocol had not yet become common knowledge, but was no longer uncommon. Someone named Dr. S. came to our rescue and placed our equipment in the data center across the street. This broker was a little smarter and understood that if we can see the entire physical data stream, then we can potentially see the orders of other clients. He was right – technically we could, but we haven’t done it yet. We had to set up a separate infrastructure for our own market data feed configuration. This broker did a great job and we were grateful for a smooth migration. The Muppetz experience was an exception. Brokers in Korea are good people.

We set up the same infrastructure in a new location and kept looking. We noticed that the speed of market data on different sites was very different. We couldn’t get a GPS signal to synchronize the local clock over the network, but Korea used the CDMA mobile phone standard. Local manufacturer Endrun Technologies sold a nice little plastic box with which we were able to get a stable time signal. This device received a GPS signal inside CDMA and transmitted NMEA data plus (over another cable) pulses at ultra-precise intervals over the serial port. The CDMA specification promises jitter for precise time signals of no more than 10 microseconds, but in fact we got about one microsecond in Seoul. In comparison, one Canadian CDMA operator showed an accuracy of about 5 or 6 microseconds.

We ended up co-located in the KRX underground data center. Getting inside was quite difficult due to strict security. Having traveled a long way underground, I noticed that a cell phone signal was being picked up inside the data center. I put our CDMA receiver in a rack and it worked – we got the exact time from the intermittent cell signal until something inside burned out. Suspecting that I needed to put in a more efficient antenna, before my next trip to Seoul, I bought a white flexible antenna a meter long and soldered an SMA connector to it. When I went down to the underground bunker of the KRX data center and put the big white antenna on our rack, the guys from Koscom looked at me very suspiciously. I don’t know what they thought, but several other people over the years have mentioned that they saw my white antenna lying on top of the rack in the KRX.

Now we had the exact time at all points of presence and we needed to establish data exchange between them. We found that a few blocks away from campus, we were getting 300 microsecond latency between our locations over the city’s ten megabit Ethernet connection. This was a very good indicator and we coordinated all of our market data and orders on these internet lines. In Korea, the Internet is wonderful and we had enough of it then. Milliseconds of benefit from receiving market data from all points of presence at once improved trading. While our best orders came from the stock exchange campus, the fastest market data came from a location a few blocks away. This turned out to be quite unexpected and that location remained the best place to get market data for several years.

A little later, we set up market data at a new point in the annex to the KRX exchange building and the delays there turned out to be huge. This was the first time I ran into Citadel in Korea, I could tell by the timestamps on the network packets – they can tell you anything if you know how to listen to them. We asked to disable this channel, Koscom stopped the transmission and physically disconnected the line. After a month or two, we paid to reconnect, they turned on the same line without changing anything, and now it was one of the best. Such were the vagaries of life in Korea. I suspect that some new kind of round-robin balancer somewhere in the middle could improve the situation. For other oddities observed there, I could not think of any obvious reason.

Virtual channels for sending X.25 orders were 64 kbps each – this one is extremely slow. I thought we could use the same hack with Sangoma cards and started writing code that would send a packet with an order before it was fully formed in memory, which would save us milliseconds. After sending the order, we could turn the package into an invalid one, if suddenly the market situation had time to change and the order being sent was no longer needed. However, our small team was already very busy, and this task was not a priority, because our trading performance was already quite good. We learned more about implied volatility modeling and fine-tuned strategies.

Only change is permanent

Exchanges never stand still and KRX was no exception. They began to abandon the X.25 protocol and moved to Ethernet and TCP/IP. The exchange planned to use TCP for sending orders and UDP for market data. The trading efficiency of our X.25 hack was starting to fade. We have become too fat with our trading engine and strategy. The performance of our trading engine hasn’t improved much from the original 300 microseconds, it’s bloated. The guys at some stage repeated my 72-hour hackathon and turned the engine into something more traditional, simple and fast. Oddly enough, we didn’t really care about a hundred microseconds here or there right now, for we were pursuing strategic and tactical goals. Compared to the rest, we actually had negative latency due to the network adapter running in promiscuous mode, modem bypass, and our own HDLC parser. Our orders were still leaving before market data was received by the rest of the market participants.

We wanted to try X.25 pre-shipping, but we didn’t have people available to spend a couple of months programming, especially since X. 25 is coming to an end. All X.25 code would be useless a few months after this project was completed. However, the idea needed to be tested, and I talked to the guy who wrote most of the software at Sangoma. He was the main developer and did the bulk of the work of writing the device driver. Maybe he can recommend someone? He said he would do it himself as a side project from home. There were some problems because the code of these drivers was very confusing with the abundance of goto statements in C, but even if this is not a very good style, the drivers worked stably. This code wasn’t supposed to last more than a few months, so it seemed like an acceptable risk. Testing took a little longer than expected and pre-submission worked, but performance was terrible. The code was so devilish that the “fast” path ran slower than the regular path. However, we did not have time to rewrite the entire code for this idea again.

Changes to TCP support on KRX turned out to be a simple mapping of X. 25 virtual circuits to TCP sessions. You can have multiple TCP sessions open at the same time (which they call PIDs) to send orders. The market data transmission format has been ported unchanged to work over UDP. I don’t remember if they added encryption to the TCP version immediately at launch or a little later, but now when sending an order, the entire data packet that comes immediately after the TCP header must be encrypted using the ARIA cipher, and this was a mandatory requirement . ARIA is the national Korean cipher. It is very similar to AES with a slightly different S-Box layout. This is a good cipher.

We saw that the new TCP-based architecture was going to shake up the market and potentially destroy much of the big brokers’ business. Good old X.25 was slow, but channel interleaving meant you had many parallel paths to send an order that could be shared. In the new world of TCP, there were a limited number of sessions, two or four, and all clients had to send their orders through these connections in turn, despite the seemingly many PIDs. As a result of this serialization, client requests queued up at the broker. The race condition now arose not only on the stock exchange, but also within each broker. Large brokers could lose their market share as the available communication lines could only support a limited number of HFT clients. In addition, the speed at which orders were sent increased, the sizes of market data packets also increased, and the rate at which market data came in did not grow proportionately to keep pace. The exchange was confidently moving towards a complete mess, and we understood that we had to somehow get ahead of this process. It was time to increase the number of our communication channels, which meant even more brokers to open accounts with.

We went to Seoul to look around and were able to get some architectural diagrams that the Koscom provider kindly attached to the brokers’ racks. There was a backbone with a bandwidth of 622 Mbps, covering the entire city and having several branches that went to brokers. In the racks we saw our old friends – T1 and E1 modems. This was good news – we could continue to use the old tricks to get market data. Our new configuration included several T1 modems that were plugged into relatively inexpensive Cisco switches. If we could listen to the T1 twisted-pair stream after the main switch decrypted the packets from the common optical channel, everything would be fine.

The transition to TCP caused a lot of confusion for KRX and everything happened exactly as we expected. Major brokers such as Samsung Securities lost their market share as they were deceived by the exchange about their ability to serve their large customer base efficiently. Market data delays became common as the exchange was now accepting orders faster without a proportional increase in seed branch performance. But we’ve already done a good job of using stock timestamps to track delays in market data. Puts, calls and futures came through different channels and usually only one (and sometimes two) was delayed. We sensed this imbalance of delays and switched to hedging mode. Several times we talked about taking advantage of the situation at those moments when we knew for sure that others were receiving highly distorted market data, but these improvements were not included in the working plans. Sometimes the imbalance of channels led to serious delays – when the data was already tens of minutes behind – we stopped everything. As soon as the imbalance disappeared, our engine automatically resumed trading.

Working with the network in promiscuous mode was interesting and did not break the rules, because I don’t think their compilers knew about it at all. Switching to TCP/IP instead of X.25 required the use of firewalls, but we learned that some local traders were working around them. It was clearly against the rules and therefore we were not ready to do this. If you break a clearly written rule, you can be completely disconnected from the KRX exchange. Moreover, she did not have a reconnection procedure explicitly prescribed in the rules if you were once kicked out, so the decision may be completely final. We calculated that our advantage was much greater than the delays on the firewall, so as long as these clients did not listen to everyone and cut packets, our advantage was out of danger and those guys were barking up the wrong tree. I recently found out about another local group trading around the firewall in Korea. So it happens, but now it’s rare.

FPGA hack

We also managed to improve our market data hack. We were a small firm with a few people named Thomas. It’s incredible, but the eighteenth hired employee became our sixth Thomas. Part of the blame for this lay with me – I knew about the “Thomas problem” and, for fun, looked for graduate students with this name in different places. I noticed the new Thomas at UNSW in their NICTA group. He was preparing for his PhD and wrote an interesting paper on distributed shared memory on Linux, including a version for Itanium processors. I had come across this before, so I made a deal with him and he started working part-time as soon as he got his PhD. One of the cool things he did was crack the Sangoma NIC firmware we were using. These cards were based on a small FPGA chip from XILINX, namely the Spartan 3 model. He accessed the flash memory in the FPGA and found some registers that we could control. Soon we had our own working firmware for it, and we were able to completely bypass the stock Sangoma firmware with its promiscuous mode. Sangoma has become the main equipment supplier for us. The discovery that the Sangoma networking stack cost us almost a full millisecond came as a shock to us. The latency of our homemade networking stack was only about ten microseconds.

Since we planned to enter other markets, we began to explore the FPGA more closely. Some time ago I almost started using FPGA instead of TopSpin’s Ethernet to InfiniBand converter. After tinkering with it for some time, I approached an FPGA company called Celoxica back in 2004 to find out about their technology. When we first started our HFT project, I asked Celoxica to come up with a solution to receive data over UDP. Trading was a new direction for them then, but they did some preliminary research and returned. Celoxica had some code for UDP over Ethernet decoding that they used in an embedded car entertainment system in Germany that could be repurposed to suit our needs. We explored it a little, but this option turned out to be too expensive. Years later, I was talking to one of the CTOs at a large Wall Street Bank who told me that he started talking to Celoxica about trading around the same time as I did. Therefore, Celoxica began development in this direction and subsequently it became an important source of income for them. Their VHDL add-on, which they called Handel-C, was quite beautiful, but, unfortunately, it could not fully meet the requirements specification. We eventually dived into FPGA development, which included the exploration of yet another language, Impulse-C. He was pretty cool, but we went back to VHDL because our design fit his model better.

Our new SANGOMA firmware worked great for us. However, just as Einstein imagined himself running next to a beam of light, we wanted to hit the first bit, an electron or a photon, arriving at the broker. This means that we must be connected to a common fiber optic cable entering the building. This fiber was connected to a Telco communication closet from which E1 modems protruded. We wanted to get right into the optical line, bypass all network devices and leave E1 copper behind. Nowadays, you can easily buy a lot of FPGA cards that already have the necessary network connectors, but then we could not find anything that could work at the 622 Mbps speed that Koscom showed.

So we designed our own PCB to get the PCIe network adapter we need. We directly brought the SFP connector to the FPGA chip from XILINX and nothing more. We have previously made some circuit boards for T1/E1 adapters which we have packaged in a small black plastic case to simplify our infrastructure. Those circuit boards only cost a dime or two each and helped save our nerves as we no longer needed active adapters from NetOptics and our cabling became simpler. Our FPGA adapter was very simple by the standards of FPGA devices, but not a simple thing for us. We brought these adapters and several NetOptics passive optical converters to Korea. After midnight, when everything was quiet, we ran into the cables and plugged our adapters into the gap. Now we had a signal, but it was completely incomprehensible.

Back in Sydney, we analyzed the mirrored bitstream and struggled to understand the signal. In the end, we guessed that the architectural scheme that hung on the racks in Korea was wrong. We actually received an STS-3/STM-1 optical carrier with a speed of 155 Mbps according to the OC3 specification and now we could decode this stream. We had more bandwidth available and were even able to find our own orders in data packets, which made us very happy. Other clients’ orders were encrypted (and that’s for the best.) Now that there were many more 64 kbps channels to scan, we had to write our own tools to find the necessary data, because everything changed too often. In the end, we ended up moving to our own star optics in a rack near the KRX campus anyway, using our FPGA adapters, which proved to be quite suitable for arbitrating and coordinating as many as eight points of presence.

Preliminary Order

Now I wanted to try that new KRX order trick. Since it was a TCP protocol, we could also immediately send part of the packet with the order to the channel until the entire incoming packet was received and decoded – these are just bytes. The secondary TCP header was fine, so we would have saved quite a lot. However, we had two problems. First, if we didn’t have an order ready to ship, what would we do then? Every N seconds we had to send a heartbeat packet if the PID was not in use. We couldn’t send a trash bag to the exchange, because then we could draw too much attention. Since we were behind a firewall, we could only work with the contents of IP packets, and no Ethernet checksum invalidation or other similar tricks at a lower level worked from here.

A cancel command for an already canceled or filled order seemed like a good message – this often happened during normal engine operation. So we sent her. Each time we randomly chose an order ID to avoid repetition. We occasionally received emails asking about these cancellations (which were sent by a smart broker), so we made a disguise – when the market was quiet, the engine would fall back to the tactic of sending heartbeats packets. All this worked fine in the unencrypted test network segment, but we still had to solve the problem with the ARIA cipher.

It’s worth noting that even on a 1Gbps link with FIX 4.2 protocol (like those used on the US BATS exchange prior to the 10G ubiquity), the savings on this pre-sending still works if you’re using an FPGA. For every 25 bytes you save 200 nanoseconds. Sending an invalid checksum of an Ethernet packet at the end will usually prevent the switch from processing the packet.

Cracking the ARIA cipher

I wrote a letter to Seoul University, which was the owner of the cipher, and they sent me the original C code for the 8 32-bit versions of the cipher. We looked at it and were able to change the handshake process to broadcast the session key to the outside and after that we were already able to encrypt messages correctly. There was some questionable aspect to the whole encryption business model as you had to pay a royalty of several thousand dollars a year to the firm to use their official open source encryption library, however everyone in Korea was paying for it.

Using our own implementation of ARIA, we have become much faster without the official library. In the end, we wrote our own SIMD version, which worked even better. Our reversed encryption allowed us to wrap our entire trading cycle in less time than the service code of the original cipher was forced to execute.

Now we could finally send the provisional order over TCP. On this, we first saved a few milliseconds, then this gain was gradually reduced to hundreds of microseconds with an increase in the speed of the channels. Our advantage was actually even more than a simple reduction in latency. The Koscom switches worked on a “buffer then send” approach, so small packets traveled through the entire system faster and the resulting improvement in response time was a little better than we expected.

There are many other open source implementations for the ARIA cipher today, and your Korean life might be easier than ours.

Shake your wings

We also tried some packet fragmentation techniques. A network device may split a packet into fragments for various reasons. The default Unix stack will hold fragments in a buffer for sixty seconds and reassemble them. If nothing else arrives, the packets will be dropped. The idea was to first send the possible tail fragments of the packet, then send the correct middle or leading fragment and let the switch reassemble the entire packet. The line speed was so slow that it could save hundreds of microseconds.

We tweaked the network stack code to test this idea. Fragmented packages didn’t work at first. However, after reducing the interval between fragments to less than two seconds, everything started up nicely. It looks like the switch treated the packet fragments as a denial of service attack or something similar if the interval was too long. He should probably not miss them at all, but who are we to argue with him? Using this approach, you could have prepared packages waiting right in the path to be merged with other fragments later. However, the two second interval was a problem as we consumed a lot of valuable bandwidth trying to keep this temporary packet cache alive. We had a better idea.

We have already sent the first part of the message in the first packet. Why not send the later part of the packet with the next TCP sequence number plus one, and then send the middle part of the sequence outside that numbering, containing the ticker, price, and quantity?
This final part of the package had fairly large custom fields, so we would save a huge amount of time and get a lot more than with the pre-submit we were already doing. There was only one big obstacle to this plan.

ARIA cipher was used in block chaining (CBC) mode. This means that the encryption of subsequent bytes depends on the contents of the previous ciphertext. Because we didn’t know the middle of the package before pre-shipping, we had a problem.

If the last bytes of the packet with the order could be random garbage, it wouldn’t matter, since decryption would turn garbage into garbage (which would be acceptable.) Unfortunately, some important fields must be filled in after the large custom field at the tail of the package. In addition, the KRX exchange was afraid of zero bytes in packets. She and her technology provider wouldn’t allow it. Koscom seemed to have an unfounded fear that something might explode or terminate prematurely if a null byte were present in the packet. However, having a bunch of zeros in the encrypted stream was allowed because they couldn’t control it. The ciphertext with zeros seemed to force a mini-reload (state reset) in the sequence of blocks due to the CBC XOR operation with one zero operand. That is, it is enough to generate data, after encryption of which a zero byte appeared, in order to interrupt the chain of dependence of blocks on each other. This generated data can result in anything after being decrypted, even illegal null bytes.

Our hack was to pass this generated data in a custom field in the batch, and also save the middle chunk with a few characters of that same field. We have to fragment the TCP packet by breaking the user field into two pieces. You have a 90% or more chance that the garbage you have in your custom field is not zero. If you check this when preparing your middle packet and see that it is not, you simply cycle through a character or two in the user field in the middle packet (i.e. the last one already delivered) until you find a completely non-zero sequence in the ciphertext. This hack can save you a lot of time.

Machine learning

The biggest improvement we’ve ever made to a strategy didn’t come from these low-level tricks. We used Random Forest (RF) to work as a machine learning layer to predict the future theoretical price. I first assigned this task to one of my PhDs at the firm (we had six. ) He seemed to be a math genius and had a PhD in physics on some kind of singular cosmological model, and was also a good friend of mine whom I hired several years ago and taught him C++. He subsequently lectured on C++ at the University of Queensland. I also hired one of his graduate students, whom he taught C++ (it’s funny that the company had several generations of C++ learners.) I tasked him with developing a machine learning layer using RF for the strategy. He came back a couple of months later and showed me all the things he tried that didn’t work and then explained to me why it would never work at all and why it was all a waste of time.

I had already come across this method in graduate school, so I decided that it was better to do everything myself again. I picked up an OpenDT v5.2 package that ran via MPI on our mini-ITX board cluster. Now it has grown to over a thousand cores with motherboards screwed to chipboard shelving, and power supplies nestled in perforated metal shelves at the top. Two weeks of running OpenDT on this cluster helped me make a decent decision. I could replicate that result using the original code from the author of the Random Forest method, Leo Briman. I threw my work over the fence to Mr. L. (the guy who sent a thousand orders per second to the exchange, which almost killed us then) to roll out into production. He was a decent programmer (albeit an introverted dude) and everything was ready in two weeks. We saw an immediate increase in profits of about a third – it was a good month.

Further improvements and adjustments to the ML layer yielded greater returns. We have become very dependent on it. Creating a fast response random forest was hard work, as accessing a huge amount of memory at random is the bane of low latency. In later projects, I did it a little better. After about a year or so, it became clear that if we turned off ML, we would not be making any money at all. However, we never really explored whether this was due to our strategy or if one low latency was no longer enough to win. I suspect the second. I’ve seen smart players like Jump set up dozens of servers in Korea and they clearly haven’t been using them to idle. There was enough ML in the air.

One mature way of thinking about ML is that it’s just a latency reduction technique. He buys you time by trading in uncertainty. You react early, but you can be wrong. However, even the best arbitrage bots do not necessarily get filled on all of their orders. It’s more correct to think of ML as a way to get additional anchor points in the space of delays and uncertainties.

Canary

Another hack that I should probably mention that is common knowledge is the canary technology. The CME was criticized by many traders some time ago, as the Wall Street Journal published a study showing that some traders were getting their orders filled before the market data that would have triggered those orders came in, and this speed advantage gave them an edge. over regular members. This is a fallacious argument, since the speeds of two such different processes can never be the same. This is rather a reason to introduce additional processing code to receive information simultaneously from the slow market data channel and the fast order execution notification channel.

There are several exchanges where the situation is different. Nasdaq was one of the first exchanges to deliver market data faster than order notifications. It was a trap for young players who didn’t expect this.

In my first experiences with the KRX, the canary would have been triggered a second before the market data. Let me remind you that a canary is a limit order that you placed in advance at the depth of the glass, and not at the best level. Thus, when the market suddenly jumps to a new price level, your canary starts to sing, letting you know about the move a second before the rest of the market. If you can’t make money from that either, your skull must be filled with stones instead of brains. Well, I found some rocks in my head.

It was strange. Sometimes, the price difference between these two channels was so big that you could see the market go up and then down. If it’s naive to buy on that first move up, the market can give you problems when it goes down. You could instead try to pre-time these events so you can buy low and sell your contracts right away.

Too much knowledge was a funny problem. We have integrated this code into our data handler. One of the guys, Mr. F., called this event “star power”, like the Mario brothers.

Jitter vs Delay

You may recall how the KRX integration of three exchanges didn’t really affect KOSPI futures and options as they were already trading in Seoul on the KSE. The united exchange KRX closed the data center in Busan and everyone moved to Seoul. However, the government did not achieve its political goals and the Busan data center eventually reopened.

Banks and brokers in Seoul spent a lot of money on these transfers. Sometimes it turned out that there was no peering between two data centers in Busan, and market data first went to Seoul and then back. Disorder. Only now, in 2016, the KRX exchange has moved back to be centered in Busan. Before that, the best solution was to collect market data in Seoul, and sending orders to Busan was expensive and inefficient. All HFT firms (like us) considered building their own lines between Seoul and Busan, however, Koscom banned this in order not to lose their revenues. You could only buy expensive lines from Koscom, which is what all brokers did. As a rule, you fought for channels that gave a delay of 2.9up to 3.1 milliseconds. Pretty slow for a 300 km straight cable or a 400 km road. Only in the last couple of years have several microwave radio relay links appeared.

At the same time, the KRX data center has evolved from the world’s largest commercial platform on HP/Compaq/Tandem hardware to an AIX platform running on the IBM PowerPC architecture. Before this migration, we were getting 11-12ms round trip (RTT) pings on order send lines with about 130ms jitter. After moving to Busan, the ping was over 20ms, but the jitter was down to 30ms. Even though latency doubled on average, reducing jitter made our strategy more efficient.

This is similar to when CME reduced jitter by replacing its gateways with iLink’s FPGA based hardware. Randomness decreases and speed becomes more important. The difference with Korea was that the delay actually increased for KRX, but as traders we don’t care, more important is the answer to another question – is my effort worth it? The new platform has improved our ROI in Korea. Similarly, I expect that less jitter, as a result of constantly reducing intra-exchange delays (not only on the CME, but on exchanges around the world), will continue to divide traders into those who can trade quickly and those who cannot. Delays continue to matter, even if they are not critical.

Competition

It’s worth mentioning what happened when Getco came to Korea. I don’t know much about them in a practical sense. The first thing I heard was constant complaints from brokers. Getco manages some pretty tough deals, and since they know how to make money from it, it worked in Korea too. There were rumors of amazing deals with the brokers themselves so that they could increase their market share. Perhaps this is not too different from the Canadian deal with CIBC? At the time, Getco was rumored to have paid a very low fixed price (perhaps as little as a few thousand dollars) for a Canadian broker to secure market share for CIBC.

I started reading reports about Getco taking 5% of the market. Someone told me that this figure was closer to 15%. Many brokers complain about their clients not getting enough commission from them. The local community was quite vocal and frustrated about the arrival of Getco. But we weren’t upset one bit, at all. We were diversified enough and traded strategies that no one else had, so very few people could even understand what was happening with us. However, it served as a lesson – Getco brought in new people by trading more effective strategies. Many people lose their income and complain, the press spins a scandal. You start reading stories about how unfair the market is. It’s not like that – just an echo from the screams of the losers. This is the sad story you hear over and over again in the markets as traders fight for market inefficiencies and the effectiveness of their technology. The market is improving, but the complaints are getting louder.

What we did worked and was acceptable in Korea, and it was acceptable in many other countries. But not, for example, in Eurex. The rules differ from market to market. If you are a portfolio manager, you now realize that you no longer want to play this game. It’s a tough competition, with hundreds of smart traders like me constantly coming up with new strategies. But you do not need to play on this field. A high frequency trader like me is next tick oriented. That’s all we care about. We’re spending huge amounts of resources pushing our fellow HFTers out the next tick. I don’t care about your portfolio alpha. It’s beyond my horizon, beyond my concern.

I lose a lot and win a lot during the day. The HFT firm described here, which I founded in May 2005, I left in January 2011. We had only eight losing days in all this time on our main trading strategy. In four of them, the loss was less than 1,000 USD and in all of them it was less than 10,000 USD. It’s not as good as the famous Virtu (1 day out of about 1200 according to their IPO paperwork), but I’ve come to terms with it. However, remember that among the transactions there are a lot of those that cannot boast of a high financial result. We only earn a small portion of the spread, usually 10% on a normal day, 5% on a bad day and 20% on a good day. Options expiration days have sometimes given us 50% as many traders try not to trade on those Fridays every month.

The law of large numbers allows HFT to make money every day. The same law states that if we lose our technological advantage, we will either not be able to trade at all or we will lose money every day. HFT companies often fail and leave. It is a tough industry, constantly delving into the infrastructure and microstructure of the market in order to improve its efficiency. And it will be a thankless job if people don’t understand that today’s rooster can instantly turn into a feather pillow when a company like Getco comes to your city. I’m sure the arrival of Jump also brought in a few dusty pillows.

Korea is a funny place, as the infrastructure of the KRX exchange is still a complete mess. There are many more ideas that can be implemented in a slow, clumsy and secret infrastructure. HFTers love clutter because they can dig through it and find advantages for themselves. It’s not that difficult, it just needs more attention to the smallest details, because any complexity consists of simple things. Oddly enough, HFTers hate exchange colocation and low latency, as it is much more difficult to find an advantage there. HFT players also love venues that few people know about… Nobody wants to be ambushed.

Portfolio Manager

Don’t worry about the next tick. Reward market structures like ETFs that are competitive, fair, and cheap. Then you can stand and rejoice as these HFT companies kill each other for the lowest spread. This allows the liquidity they provide to work for you, regardless of their motives. After all, only buying higher and selling lower (ie narrowing the real spread) works in competition for the next tick. This is not a recipe for easy money in HFT.

HFTs are afraid of portfolio managers – they wield a big information club that kills HFTs. Large orders move prices, which prevents the market maker from earning his spread. Asset managers even think differently. Each order execution is a desired trade, even if the market has gone further against you and the account has a drawdown, and the loss is simply an expense that is always within your model. In HFT, execution is a risk, a drawdown is a disaster, and a loss is just a shrug.

A portfolio manager must use algorithms and the magic of market efficiency to improve the quality of their order execution. You don’t want to hire twenty people at once just to compete for a position in the queue on the next tick? You should be more afraid of machine learning-based quantum funds that are steadily automating your alpha. Worry about frustrated HFT traders who decide to change models and invest long term to compete directly with you. I see that both are already happening.

Very few brokers offer competitive solutions, but for the most part everything seems a bit stale. A portfolio manager of a large fund with a large team can make additional profits by deploying their own algorithms and DMA platform so that their traders can trade more aggressively and frequently, in the same way that floundering HFTs look at their time horizon. But beware – it’s a slippery slope…

Previously published under the pseudonym Thomas Anderson. This text omits the sex, drugs, and rock and roll from the original draft, as it would add nothing to the trading story. Perhaps this will help justify my decision to turn down a job I enjoyed. In any case, all this is best left unsaid, even if some personal relationship is missed.

The deal for De Gea and Navas failed in the last seconds, Manchester United buys Martial for 80 million, Kokorin was not released to Arsenal.

Online closing of the transfer window

31.08.2015, 10:00
53

Today is X day for most European clubs. FootballHD.ru throughout the day will follow the loudest transfers. Stay with us!

02.25 All this, of course, is insanely interesting and funny, but it would be time to be honored to know. All other information – tomorrow in the new online on the Premier League. Thank you for being with us and see you soon!

02.24 We decided that UEFA would make the decision tomorrow. How many more details are waiting for us?

02.17 And more details. The documents were sent with a delay of 20 seconds! Twenty! But the deadline is made by a machine, not a person. And she has no heart. There is also a version that the turmoil occurred for a very interesting reason: United employees sent documents in a format that was not opened either in the Blancos office or in the Spanish football federation. Maybe Enigma encrypted it?

02. 14 More trash! The Guardian journalist Daniel Taylor breaks into our online and claims that, firstly, there is a special mark on the Manchester United documents, which indicates that everything was done on time, and therefore the Navas transition was completed; secondly: “Real” is furious and intends to seek a review of the decision. It’s understandable, “United” in chocolate with three great goalkeepers, and “Real” with Casilla. This is karma.

02.10 Colleagues gradually finish online and go to sleep. But how can you sleep here when there is so much fun around? Therefore, we will sit for a while, and then we will disperse.

02.04 Clubs still blame each other and can’t figure out what to do about it all. “Madrid” assures that they have proof that everything was done on time. But there is an opinion that instead of papers, a fax came to the office of Real Madrid with the following content:

02.00 And who are our kings of scams?

01. 57 Ricky van Wolfswinkel , aka PigWolf, went straight from the Norwich Championships to the Provincial Betis. And he was interested in “Spartak”, which is not difficult to see the irony.

01.52 Manchester and Madrid employees begin to swear and exchange complaints. And on Woodward they already draw such photoshops.

01.46 Ahaha! Everything is more fun and more fun with us. It turns out that the deal was executed late for one reason – at the last moment, changes were made to one of the clauses of the contract. Guess who. Keylor Navas! This is how you win the competition!

01.44 In general, we have complete surREALism. Let’s wait for news until victory. We don’t split up!

01.42 OU! AND NOW THERE IS INFORMATION THAT MU SENT THE PAPERS AT 23:59! Remembered the existence of bribes and decided to replay everything?

01.41 Miguel Layoun has moved from Watford to Porto. You may remember him for his incendiary performances for the Mexican national team at the World Cup. Will now close the gap left by the departure of Alex Sandro.

01.36 In the confusion, we completely forgot about Yuri Zhirkov . He almost moved to Zenit, but rejected the contract offered to him. Money not arranged or something else? But it doesn’t matter anymore. Yuri remains in Dynamo.

01.32 No, do you have any idea what it is? Two giants were left with unmotivated and frustrated main goalkeepers. De Gea had already signed a contract and was ready for the presentation, and Navas was kicked out as unnecessary. I wonder who is meaner now? And which club will be more affected by this situation?

01.27 HAHAHAHAHAHA! I CAN `T BELIEVE MY EYES! DE GEA’S TRANSFER FAILED! Paradox! Just a paradox! They procrastinated all summer, and on you! We agreed on everything, but in Manchester the fax was sent too late. “MJ” can still buy Navas, but now it is unlikely that “Real” will sell him. “Madrid” does not have the right to sign anyone before winter. Yeah. Such huge clubs, but like small children.

01.25 There is information that Arsenal is interested in buying Cavani . How timely, huh!

01.23 YOYOYOY! SPANISH MEDIA CLAIMS THAT “REAL” DID NOT HAVE TIME TO MAKE DE JEA’S TRANSFER! Awaiting confirmation.

01.21 Sadness-sorrow. “Marseille” was left without Lamela . Tottenham blocked the transfer. And there was no time for negotiations. The French were not allowed to shush.

01.18 News from Rome continues to come. Roma signed Ezequiel Ponce on the flag. Another great talent from Newell’s Old Boys. Attack. It cost 4.2 million. Contract for five years. Doumbia is clearly not going back.

01.12 And we have a new sensation here! Rojo quarreled with van Gaal and threatens to leave United tomorrow !

01. 08 Visual graphics of Manchester United transfers. But they lied about Martial.

01.05 And what about De Gea?

01.04 Boo! The window was covered everywhere except England .

00.59

Yes.

00.57 Breaking the stereotype about Porto.

00.55 Tottenham won’t give Lamela to Marseille until they find a replacement. Well, greedy! You have half the team playing in this position there!

00.53 Ledesma from Lazio played enough in Europe and decided to return to South America. And not just anywhere, but in Santos.

00.51 Kemen left Newcastle. Where? That’s right – in Marseille.

00.49 And according to Canal+, Martial is given for 50, plus 30 as bonuses. And not a word about ZM.

00.43 Siqueira stays with Atlético . Benfica in flight. And almost snatched it from Juventus. Now we have to wait until winter.

00.41 Aston Villa want to sign Lescott and Adebayor . But it doesn’t seem to be today.

00.39 And there are no more deals in Russia and there won’t be any more – the league leadership has gone home. The day was quiet and boring. One word – limit.

00.37 And the Nerazzurri rented Alex Telles from Galatasaray. Also Brazilian. When are they all done?

00.35 Inter have confirmed the signing of Felippe Melo. For 3.5 million, as they said. So we’ll find out if the simulators are lying to us or not – is he still in the top 50 players in the world or have they gone crazy in Milan?

00.31 Nope, not 80. The contract is crafty. United pays 40. And will pay the same if Martial wins the Ballon d’Or. All management will now root for Messi and Cristiano.

00.27 Kanye West is running for President of the United States, Martial is being bought for 80 million, Rennes are loaning a black God from Chelsea. .. Am I falling asleep in the med school dorm and having a nightmare?

00.24 Less than an hour before the window closes in Spain. And where is De Gea in the new sweater?

00.22

And this is Bruno Soriano . Here is what Nobel Arustamyan said:

″… Soriano did not have time to sign all the pages of the image contract with Napoli. The fact is that the president of Napoli, De Laurentiis, makes every player sign an image contract as well. And there are 46 pages …″

This is about the benefits of excessive meticulousness.

00.19 А The transfer window for has closed in Russia. Waiting for another sensation after the fact?

00.17 Louis van Gaal clearly outplayed the fourteenth fifa. Otherwise, I can’t explain it. Is there anyone alive here? Do you think the Frenchman will pay for such a price tag?

00.10 TAAAAAAAAAAAAAAAAA! CLOSED ONE OF THE MAIN QUESTIONS! MARTIAL TO MU! 80 MILLION! EIGHTY! EIGHTY!

00.08 No, the news from Italy will never end! Ljaich has officially moved to Inter ! Like they said, lease with option to buy.

00.06 Nobel says he did it. Although he told us about Klopp today.

00.04 And yet there is one intrigue from Serie A: did Soriano manage to sign a contract?

00.02 Oh, and summer is over!

00.00 It’s midnight. A fabulous time of realizing that there is definitely nothing to expect from Italy. And who will now slow down the transfer market? Marseille alone?

23.58 Wow! Witsel remains ! I don’t know whether to rejoice in this or not, but this is what Galliani said:

″. .. We tried to transfer Witzel, but we did not succeed. We knew from the start that this was how it would end. …″

Why did you get involved then?

23.56 It turned out that Bibras Natkho was against his brother’s transfer to CSKA. Like, you have to stay in Barcelona. It seems that someone will become less likely to get into the squad.

23.52 Huh! “Marseilles”! What are you doing, stop it! The French signed Isla and De Ceglie from Juve. The first was rented and the second was bought. It looks like someone will seriously compete with PSG.

23.45 Guys, I saw God! See you too!

And the casting director for “Bruce Almighty” definitely knew something.

23.42 Celtic are being taken apart piece by piece. Forward Stefan Schepovich will now delight many Getafe fans with his goals.

23.38

Whoa-whoa-whoa! Quiet! The guys from Dynamo dispersed in earnest. In my opinion, this is an application for the championship. It is possible to sell Kokorin at such a pace, otherwise he will fall into the reserve.

23.33 Hernanes is at Juve, Hernanes is happy. Will the Bianconeri be just as happy at the end of the season?

23.30 Southampton remembered they needed centre-backs and renewed interest in Van Dijk . Not to an artist, but to a Celtic player who has openly stated that he intends to leave the club. The Saints are confident that 11.5 million will be enough to convince the Scots.

23.26 Almost half an hour has passed since the transfer window in Italy closed, and there is no information on Witsel. Looks like it remains to fight for the Champions League.

23.24 Dynamo had a bad day and summer in general. They did not buy Ryazantsev, they did not sell Kokorin, they lost half the staff. But for some reason they left Stanislav Stanislavovich Cherchesov. And they declared him to the Premier League under the norm number 43. All right, tremble, Gabulov!

23.21 RMC, who sent Martial to Manchester United today for 50 million, claims that Marseille agreed to lease Lamela . It seems to be worth believing.

23.19 Jesúsa Corona was courted by PSV for a long time, but could not win. But Porto could. He always can. And soon, for sure, it will sell dearly. Still – such a name for a person. Jokes aside, the winger is great and who the hell knows why the big clubs haven’t noticed him yet. And the Portuguese did not spare 12 million, for example.

23.17 Victor Ibarbo , who failed at Roma, the rumors are sent to Watford. How cruel and strange European football has become. One successful season – and 50 million are paid for you, unsuccessful several months – and outsiders are interested in you.

23.14 While we’re quiet, I’ve done a transfer of damn good coffee from the kitchen straight to my stomach. The calm, as you know, is always before the storm, so you can’t sleep. And don’t you dare!

23.04 Yuny Dmitry Belonog moved from Shakhtar to Ural. There he risks becoming a Mudfoot, because there is something to get dirty in. If you understand what I mean…

23.02 There is some interesting news from London though. No, not from Arsenal. It’s just that Tottenham decided to raise the offer by Beraino up to 25 million. There, unlike some, they understand that attackers are needed.

23.00 Bang! This window closed in hyperactive Italy . However, it is possible that news from it will still come. More somehow and nowhere.

22.56 Everything! Rolling up! In “Dynamo” killed the main intrigue of the day. Kokorin remains . The irony is that this would be Arsenal’s only transfer on today’s hot day. The Londoners planned to take Alexander on loan with an option to buy. But shish you, Mr. Wenger! Sasha is too good for your nuclear submarine.

22.53 At the beginning of the month, there was a lot of talk about the transfer of Mayuki to Spartak, but Dmitry Alenichev did not need him. Like Southampton. However, Emmanuel has found a home in France – now he is a Metz player. “Only not everyone can afford it. But Petersburgers… They could very well. Would you like to see how Jurgen teaches Dziuba and Yusupov sharp counterattacks with touchdown passes?

22.45 Bologna have loaned Giaccherini from Sunderland. If you still remember about its existence, then you are well done.

22.43

I join. It seems that everything goes to the fact that for any opponent from the Premier League match with “MS” will turn into masochism.

22.38 Wow! Lamela seems to be in Marseille! At least everyone is actively trumpeting it. Oh, how early Bielsa left!

22.36 Proof Grosskreuz from Turkey. And there he is not on vacation at all. Hmm, but usually, on the contrary, the Turks go to Germany to work.

22.33 Italy does not let us go. Roberto Soriano moved to Napoli for 13.5 million euros. Also included in the deal was a one-year lease for Zunigi . In the Apennines, they decided to try for the whole of Europe in the last hours?

22.30 “Carpi” did a good job on the last day. Another eminent Italian joined Zaccardo – Marco Boriello . “Biancorossi” in the “series” seriously and for a long time?

22.27 The transfer evening was flooded with rentals. And what about without their main lover? Alessandro Matri went to Lazio. He hasn’t been there yet. And I keep thinking, how can I go around all of Italy? It was necessary to become textured and not very technical striker.

22.24 Do you remember the steamy defender Doria ? Even Russian clubs were interested in him, but in the end he turned out to be Marseille. Where did not take root. And left now “Granada”. Rent with the right to further purchase for 7 million. And let him allocate at least a few thousand for a club photographer’s camera.

22.20 We completely forgot about Sevilla! And she took and borrowed defender Andreolli from Inter. Only for a year so far, and then we’ll see. A noble company in Andalusia is going, I must say.

22.17 Details about Klopp’s move to St. Petersburg are being clarified . They say that the “Zenith” deal will cost 15 million. Well, at least there is no limit on coaches.

22.15 But the Turks are not appeased. Galatasaray on loan from Manchester City0184 Jason Denyer . His name has been buzzing around lately. For those who don’t know, it looks like this.

22.08 Ageless Cristian Zaccardo moved from AC Milan to Carpi. Well, now Serie A newcomers can afford world champions.

22.03 A dream come true – Joaquín is back at Betis. Oh, nostalgia! Back in FIFA 2005, he made me pretty nervous in matches with, as it seemed, an unknown team.

22.01 Fener seems to have serious plans for the season. They also almost bought the future star of the Uruguay national team. We are talking about Diego Roland . Only the ban of Willy Sagnol – the coach of “Bordeaux” – calmed down the angry Turks.

21.57 The deal for Idea is formalized. Now he is an Olympiacos player. Perhaps congratulations. Will still play in the Champions League.

21.52 It’s time to return to the Russian land. And here everything moves in a spiral, driving us into Nietzschean moods. Pavel Yakovlev again goes on a one-year loan to Wings…

21. 47 Lucas Piazon continues to roll around on loan. This time – “Reading”. I wonder if the talented midfielder will ever show up at Chelsea?

21.45 Meanwhile, Bournemouth continues to strengthen. Managed to fish out Glen Murray who fell out of favor at Palace. The transaction amount is 5 million. One-fifth De Gea.

21.42 Louis, it’s time to spend. Or will the transfer of Martial now sharply drive the club into a loss?

21.37 Schoolchildren who have not yet fallen asleep, it’s time to practice. Simple arithmetic. The total deal for De Gea , according to James Ducker of The Times, is 40 million. Keylor Navas has agreed to be part of this outrage. It was valued at 15 million. Help Dasha find out how much the most coveted goalkeeper of the passing summer is worth.

21.33 Braun Ideye , who escaped from Dynamo Kyiv, did not really succeed in the Premier League. And now Rondon has arrived. Therefore, he is forced to bow to Greece – to Olympiacos. So far, only at the level of conversations and dreams. But, Sky Sports assures that everything is serious.

21.31 Here MEDIASET announces that Zenit and AC Milan have agreed a deal for Witsel . We’ve heard something like this before, haven’t we? A hundred times in a summer. Do we believe in this one?

21.26 There is almost no news from the Netherlands today. And here they are! And very interesting ones too. PSV set out to return Leroy Fehr to his homeland. Do not vegetate as he is in the CPR, in fact. There are still a couple of hours to close the deal. And the Eindhovenians should hurry up, because, judging by the start of the championship, they clearly need a quality defensive midfielder.

21.21 It’s done! What we’ve been waiting for all day! Corporate insider from Nobel! Did someone send him money on the phone at last?

21. 18 Lindegard officially moved to West Brom. The agreement was signed for two years. And the Dane is already confessing with might and main how he fell in love with the game of “thrushes” from his first days in England. We willingly believe. Himself from childhood for the WBA.

21.15 Looks like Totti and Garcia’s cold war has begun. Quite suddenly, the Romans set out to sell Lajicha . The talented Serb claims “Inter”. We’re talking about a two-million lease followed by a nine-million buyout option.

21.05 Sensation! “Rennes” acquired God! It is clear that money can do anything now, but so far we are talking only about the youngster from Chelsea – Jeremy God . The French took a talented boy on loan for one year.

While all the students are going to bed getting ready for the big day, we continue to close the transfer window online. I take over the night shift – Nikolai Samofalov. Well, the clock is ticking, and time is now inversely proportional to the nervousness of breeders and footballers. There are only a few hours left to pull off a couple of swanky deals and polish some of the ones mentioned below. Everyone except the English. We’ll talk about them tomorrow.

20.57 In Italy, the window closes late at night, but there, too, the main transfers are coming to an end – in austerity mode. For example, Fiorentina will rent Blaszczykowski for one season for one million euros, with the possibility of subsequent redemption only next summer for 6 million euros.

20.51 It seems that the main summer saga is coming to an end. De Gea flies to Madrid, in the opposite direction to Manchester – Keylor Navas and 35 million euros. We are waiting for the official announcement.

20.47 Juventus cheerfully hold the defense against Napoli, who offered 20 million euros for Daniele Rugani . But Turintsy need such a defender.

20.40 But this is just an atomic bomb from the Italian press! Sky Italy and Gianluca Di Marzio report that Zenit have agreed with Jurgen Klopp !

20. 35 Robbie Kruse also couldn’t stand the competition and left Bayer for a simpler club. More specifically, at “Stuttgart”, for one year lease.

20.23 Continuing the Football Manager theme. Rennes loaned Colombian Juan Fernando Quintero from Porto. 1 million plus 19 million at the end of the season if the French decide to keep the player for a longer period.

20.06 Wait! Bayer announced the purchase of Mexican striker Chicharito Hernandez . The transaction amount is approximately 12 million euros.

20.03 Porto’s Spanish striker Adrian Lopez returns to his homeland. The striker will spend next season at Villarreal.

19.59 Khodzhaniyazov from Zenit goes to Danish Aarhus , but in case of a good game he still has a chance to come back within the next two years.

19.56 We return to the Russian transfer market, and there Amkar signed the 22-year-old Norwegian forward Chuma Anene , who until recently played for Rabotnichki. Let’s see how exactly Gadzhi Gadzhiev adapts it to his schemes.

19.43 Rolando still managed to endure until the end of the contract at Porto and calmly moved to Marseille as a free agent.

19.40 Urgent! Liverpool bought a striker! Meet Taiwo Awoniyi , he is 18 years old, he is from Nigeria and he will play next season at “Frankfurt” on loan. Powerful.

19.35 AC Milan is offering a lease agreement for Witzel for 3 million. And a mandatory buyout for 22 million if the player plays 20 matches for his new club. Slightly it seems that the Italians do not quite understand who they are dealing with.

19.30 The Orlando Pirates signed 23-year-old Ghanaian guard James Akas ! He is 23 years old. 23. Yes, 23, stop rubbing your eyes. Or 63, and he is generally a coach. We’ll let you know the details later.

19.18 Genoa confirmed the transfer of Christian Ansaldi from Zenit. Rent with option to buy. The power of the limit in action.

19.14 Meanwhile, the “soldiers” officially announced the arrival of Amir Natkho . The agreement was signed under the 1+3 scheme. I wonder how it is?

19.09 So Aaron Hunt has officially moved to Hamburg. Of all the top German transfers, the only thing left to wait for is Chicharito’s announcement at Bayer Leverkusen.

19.00 Germany has completed its transfer frenzy! We are still waiting for the announcement of the transfer of some players, but by this time all contracts must be signed.

18.53 Schalke’s twitter announces that Draxler has officially moved to Wolfsburg ! The club’s website does not specify a specific amount, but “this is the most profitable transfer in the history of the club. ” But the player signed a contract for five years. I wonder how the fans of the London “Arsenal” feel now?

18.50 Only ten minutes left in Germany! The website of Borussia Dortmund confirms the new player’s one-year loan, and Januzaj has already fully revealed all the cards.

18.44 But CSKA never says goodbye to their talents – Dmitry Efremov flies to the Czech Republic for practice in Liberec “Slovan”. Not a bad choice, a loan to this club once helped Dynamo Kyiv turn Serhiy Rybalka into a key player in the team.

18.35 Fenerbahce at the finish line in a deal for Gabriel Barbosa of Santos’ . One of the main talents in Football Manager 15 has already managed to say goodbye to the fans of the Brazilian team in all social networks.

18.30 “Zenith” was so impressed with the game “Krylya Sovetov” that sends Milan Rodich to Samara to help. This is a complete transfer.

18.26 Brazilian defender Felipe joined Udinese as a free agent. According to the official website of the Friulian club, the 31-year-old defender has signed a two-year contract. The player has already played in Udine from 2003 to 2010 and this can be said to be a return to his native land.

18.20 This was already known yesterday, but the deal was made only a few minutes ago. Dante goes to Wolfsburg for 5 million euros! Recall that for the same amount he left the Mönchengladbach Borussia for the sake of Bayern three years ago.

18.10 Meanwhile, Everton fans are trolling Jose Mourinho by singing the legendary Beatles song “Can’t buy me love” in a new way. “Money won’t buy you the Stones!”

18.00 Only one hour left before the transfer window closes in Germany! We are waiting for the finalization of key deals – Draxler to Wolfsburg, Januzaj to Borussia Dortmund and Chicharito to Bayer.

17.52 Fabio Borini is officially back at Sunderland! The striker has decided not to fight for a place in the first team of Liverpool, and the cats will pay about ten million pounds for him. That is, they will completely beat off the purchase price of the player by Liverpool from Roma.

17.45 Alessandro Nesta is the new head coach of FC Miami. How quickly time flies… We read this news, remember it, and are not surprised if in a couple of years the legendary defender in the past will lead Milan. There, such twists of fate are simply adored.

17.40 A little bit about pressing matters from Lithuania. Forward of the Lithuanian “Atlantas” Andrey Panyukov , who scored twenty goals in twenty championship games, goes to “Ajaccio”. True, so far only on loan.

17.30 Did you miss the news about De Gea ? And here they are! Marca announces agreement in principle between Madrid and Manchester United on the amount of the deal! 30-35 million euros will cost one of the best goalkeepers in modern football. In parallel, there are talks about the purchase by the British of Keylor Navas , who previously refused to participate in the De Gea deal. But as a separate transfer – why not?

17.26 Meanwhile, RAC1 informs that Amir Natkho is still moving from Barcelona-B to CSKA . Most likely, Muscovites will not pay a cent to the old club in this deal. But the player was offered a 4-year contract.

17.20 And yet Schalke surrendered! Julian Draxler has already arrived in Wolfsburg, where he is shown to his new home. But about 40 million euros are sent to Gelsenkirchen, which the club will be able to spend in the next transfer window. Draxler’s agent has confirmed the transfer and a five-year contract will be signed.

17.12 Borussia Dortmund does not spare the feelings of its fans either. Following Grosskreutz, another legend leaves the team – Cuba Blaszczykowski said goodbye to the city on his Facebook. He leaves Dortmund for Fiorentina .

17.09 Russian clubs, despite the later closing of the transfer window, are trying to solve their problems now. For example, Moscow “Spartak” unexpectedly popped up with an offer for “Tottenham” for Adebayora . Bomb? And then! 5 million euros are at stake.

17.06 Can we already congratulate the fans of Dynamo Kyiv? It is unlikely that Yarmolenko will go anywhere, moreover, the club’s management still managed to shove Mbokani – Norwich needed a powerful striker and the Canaries are willing to shell out six million euros for him. The Congolese had already flown to England in a private plane, things had already been collected in advance.

16.53 Against this background, it is rather strange to look at the passivity of English clubs. Is that the saga of Martial and convulsive attempts Chelsea buy at least some player in the defensive line. But not the Stones, but Papi Djilobodji , who will come from Nantes for 2.7 million euros. The player’s medical examination is already underway, while we are still trying to calculate how much Abramovich saved on this transfer.

4.50 pm Come on? According to “Soviet Sport”, “Rubin” wants to return Berdyev . And with him to return and those players, “Rostov”, who used to play for Kazan?

16.45 Leverkusen Bayer is going to compete with Dortmund. Kicker is talking about 11 million euros for Chicharito – no loan, just a buyout of a player who played for Madrid all last season. However, Bayer will not overflow, because Robbie Kruse is going to shelter Zonninger and the entire Stuttgart in his team.

16.40 In Germany, all the main news comes from Dortmund. Bild confidently reports that Adnan Januzaj will be Borussia ‘s on loan next year. Judging by the mood of the leaders of Dortmund, this lease is with the right to purchase. At the same time, Kevin Grosskreutz leaves his hometown to join Galatasaray’s powerful roster . And for starters, Oliver Kirch will be the last purchase of Paderborn .