Furious bulls: how Wall Street became dependent on "high-speed" trades. Part 3

High-frequency traders make money out of thin air by collecting pennies that appear and disappear like virtual particles from quantum field theory. Their goal is to end each trading day without open positions, i.e. exit the market, keeping the profits in the bank. Depending on the model chosen, they may well exist, making a profit with only 55% of transactions. They constantly check prices, look for patterns and trends or the opportunity to buy something in one place for $ 1, and sell somewhere else for $ 1.01 or $ 1,001.

[ First part ], [ Second part ]
')
Sometimes they don't even try to make money on the trade itself. Under the maker-taker model, some exchanges offer small bonus payments or price discounts for requesting quotes (for buying or selling shares), followed by a trading operation. The exchange charges from the opposite side, the taker, a slightly higher fee and keeps the difference. Due to this, the algorithm can buy a share, get a discount, then sell the share and get a discount again.

All this is controlled by algorithms, the life cycle of which can be no longer than a few weeks. Sometimes the algorithm simply looks for stocks, the movement of market prices of which is intensified over several transactions in a row. Such an algorithm (momentum algo) will acquire a share, in the expectation that its price will continue to grow. The “regression” algorithm will sell, expecting the price to roll back to the average value. The same company can use both of these algorithms simultaneously. Within a minute, both the one and the other algorithms may be right.

One of the strategies common to algorithms is the search for pairs of stocks whose prices historically correlate with each other. A classic example is oil stocks rising with commodity prices, or airline stocks showing an inverse relationship. But not all of them move at the same time, so there are strategies that involve the purchase or sale of that stock, the movement of which repeats the dynamics of another security, in order to then realize it profitably.

Similarly, “derivative” financial instruments, such as options and futures, can go out of balance with related stocks. Some algorithms are “market makers” of the exchange - they are trying to buy shares at the minimum purchase price and quickly sell them at a slightly higher selling price, collecting the difference, called the spread. People who did this were formerly called experts, and earned good money when spreads reached the eighth part of a dollar. Since the New York Stock Exchange in 2001 declared “decimalization” [ from English decimal, “tenth” - approx. translator ], the size of the spread dropped to a penny or two, which means that in order to earn the same money, you now need to process many more transactions and do it several times faster. This area of ​​trading on the stock market has ceased to be a lot of people.

Stock quotes are always fixed to a penny, but real prices may vary in the range of the sixth decimal place. Not a single normal person will trade up to a ten-thousandth cent. But computers are not afraid.

A more particular example: the simplified algorithm that Mani Mahjouri, chief investment officer at Tradeworx, presented at the Battle of Quants. His hypothetical algorithm buys and sells SPY-depositary receipts reflecting the productivity of the S & P 500 index, based on the value of the 500 largest public companies in the United States. SPY depositary receipts are traded on various exchanges, including the Nasdaq in New York, but Chicago hosts the S & P 500 futures market, that is, contracts, prices of which reflect the index values ​​for weeks and months in advance.

Prices for SPY and futures contracts are usually correlated, but do not go "leg in step" - usually futures contracts are ahead of SPY by a few milliseconds. The reason for this is irrelevant, all that matters is that this rule has been observed for a long time, so the algorithm can make a profit using S & P futures fluctuations to predict the behavior of SPY for the next few milliseconds. How much profit can be derived from this depends on how quickly the algorithm can get the data transmitted from Chicago to New York. As an experiment, Majouri applied the algorithm to the full list of trades for the whole day and calculated that under optimal conditions - by which he meant data and commands at the speed of light - the algorithm could perform approximately 64,000 trading operations with an average revenue of $ 0, 0001 per share.

The specified model assumed the implementation of one share per trading operation; As a rule, high-frequency traders work with relatively small volumes, several hundred shares per unit of time, since large lots change the state of the market.

Strictly as an illustration of the model, let us assume that during the average trading day, about 150 million SPY shares pass from hand to hand. Tradeworx claims that it works with about 4% of all SPY trades, so extrapolating this data, we can conclude that it is about 6 million SPY shares multiplied by $ 0.0001, which results in ... $ 600. By itself, this is, of course, an inefficient way to get rich. But multiply the numbers by the number of similar algorithms that work at any time - and the situation will become much more interesting. One of the reasons why high-frequency trading exists at all is that these operations occur too quickly for a person to interfere with this.

Here's what you can try: open the Yahoo Finance page during trading hours on the stock exchange and enter the ticker for the quotations, say, Intel (INTL). On the quotes page, click on the OrderBook button. You will see that under the “Bid” (purchase price) there is a list of five values ​​in descending order, and a similar list, ranked in ascending order, is called “Ask” (selling price). These prices reflect offers to buy or sell a certain number of shares. Prices will be close enough to each other by value, and the numbers from the first term should correlate with the current price of the ticker, reflecting the most recent trading operation.

This is how the capital market works, although this, among other things, is, of course, a fiction. If you, as a retail client, want to buy Intel stock, your order will most likely be executed at or close to the prevailing market price by your broker based on his in-house inventory. Or, for a small fee, it can be transferred to a dealer - such an “internal player” who sells and buys shares for himself, at prices no worse than in the external market. [ In this case, the client's request will not be placed on the stock exchange, but executed inside the broker according to the principle of the best execution (best execution), so to speak, internalized - approx. translator ].

You may have heard of Bernard Madoff; this is exactly what he did in his legitimate business. Purchase and sale prices are always fixed to a penny, but the prices of real trades can be calculated to three, four or even six decimal places. No sensible trader will spend his time bargaining for a ten-thousandth cent, but computers are not afraid.

Quotes in Yahoo’s order book are probably matched by an algorithm, and you absolutely can’t trade at these prices. Even if you had access to the stock exchange, which you, of course, do not, they would most likely have become outdated long before you entered the market - either trade operations or, more likely, their values ​​would have changed even before performing any operations with shares. And this is where the real danger lurks. The point is not that people are less and less involved in the bidding process - the fact is that people simply cannot take part in the bidding.

“By the time an ordinary investor sees a quote, he, roughly speaking, sees the light of a star flying 50,000 light years before him,” said Sal Arnuk, a partner at Themis Trading and co-author of a book called Broken Markets ", criticizing high-frequency trading.

According to some estimates, 90% of requests for quotations on the main stock exchanges are canceled before trading. Many of them will never entail trading operations — they are placed to test the market, to confuse competitor algorithms, or to slow down stock trading, “clogging up” the system with queries — a practice known as quoting cheating. It may even be another action, in case traders who work with it are served by the same server. On the Internet, this practice is known as DOS-attack, which is a crime. Among quanta, this is considered as the extreme manifestation of a bad taste.

This situation has led to the emergence of a unified system of quotations of exchange quotes, which aggregates quotation requests from 12 US stock exchanges and sends them to traders - this is the nervous system of the market, feeling the brunt of the situation.

“Every time they increase capacity, the volume of messages increases just enough to make up the difference,” says Eric Scott Hunsader, CEO of Nanex, which aggregates and analyzes market data. “Now you need a connection with bandwidth to gigabits and higher - and in 2000 I could monitor the markets using only a 56-kilobit modem. You can open a quote request for free, but then we all have to pay for it. ”

To be continued...

PS The II All-Russian Conference on Algorithmic Trading will take place very soon. If you are interested in this topic - come, it will be interesting.