The presence of high-frequency traders has changed the nature of trading in small sizes and in many ways has undermined the reasons why institutions embraced algorithmic trading in the first place. As a result, more and more buy-side firms are likely to avoid algos altogether and seek block liquidity instead.

There is a flaw that underlies the premise of algorithmic trading of institutional orders.

When first introduced to institutional equity traders, the idea of taking large orders and splitting them into smaller pieces made a lot of sense. But that was many years ago now. Before the introduction of algos more than a decade ago, there was a time when small orders and large orders hardly interacted. Small orders were routed to the NYSE DOT system, or found their way to Nasdaq wholesalers. Large orders were put together upstairs on the telephone, or in the Instinet machine.

The impetus for change came in the late ’90s, when new sources of small order flow exploded onto the scene – the rise of Internet brokers, the proliferation of day-traders, and layoff business from options market-makers and program traders. Not to mention, the buy-side embrace of order management technology, the introduction of the FIX protocol, the Manning Rule, Limit Order Display, and the Order Handling Rules.

Among larger buy-side firms, the early adopters began using broker technology to slice their large orders into small pieces that were indistinguishable from the general flow of small orders. They found new sources of liquidity and lowered their transaction costs.

That was how it started: Large orders put on small-order costumes and masks and attended the small-order ball. Later adopters found there were fewer and fewer revelers back at the large-order mixer. Old school traders found it hard to resist slicing large orders with algorithms, not only to interact with the huge wave of small orders but also because that’s where the other large orders were. Eventually, they too put on small-order masks and forsook their market for the Algo Ball. And this trend continued to grow because it was efficient and lowered trading costs.

And it worked very well for a long while; but it doesn’t work so well anymore. It doesn’t work because the presence of large orders masquerading as small orders has attracted a new kind of small order to the ball. High-frequency orders simply weren’t there when the algo revolution started. But they’re there now, and they aren’t leaving. Their presence has changed the nature of trading in small sizes and in many ways has undermined the reasons why institutions embraced algorithmic trading in the first place. And those masks aren’t fooling anyone anymore.

Maybe it’s time for buy-side traders to take off the masks and go back to their own party.

Set aside the metaphor and let’s talk about what’s really going on in the markets. The term we use for “not fooling anyone anymore” is “information leakage.” The traditional kind of leakage occurred when your broker shared information about your order with people he shouldn’t. Leakage nowadays is different. Who you tell – or who your routing technology tells – is still important. But in the high-frequency age, the most pernicious cause of leakage is predictability. Algos have specific objectives that make them inherently predictable. Hence the problem.

To fight predictability, algorithms employ strategies that try not to be so obvious. They randomly generate displayed quantities, and use randomized time intervals between order replenishment. They randomize trigger sizes and randomize distribution among venues. The problem is these random values are all generated between parameters that are set by the objective of the algo. Over time these parameters can be inferred because the algos generate so much data. Large orders sliced into small pieces create many orders, many trades, many data points – and many valid samples.

This is not leakage caused by dark pool IOIs or conditional order technology. It doesn’t happen because a counterparty fades on you. It’s not based on who you trade with. It’s caused by your trading. Even an algo that avoids HFT counterparties still leaves a trail of data.

And your data gets analyzed, make no mistake. Take VPIN, which is highlighted in an article by Cornell professor Maureen O’Hara in the current issue of Financial Analysts Journal. VPIN, which stands for “volume synchronized probability of informed trading,” is a real-time measure that seeks to capture the probability that somewhere out there is an informed trader. Or a large order. You have to ask yourself: If this is what’s being published, can there be any doubt that even scarier stuff is already deployed against you in the marketplace?

Going forward, traders will minimize this forensic leakage by adopting strategies with broader parameters, less predictability and fewer trade executions printed to the tape. There are multiple ways to measure this, any of which might emerge as a standard. Average trade size doesn’t meet the requirement, as it obscures real differences among stocks and order sizes. But firms might choose to track the number of separate executions per completed order, or per hour, or per 100,000 shares traded. You could measure how many executions per million dollars traded, or per percentage point of ADV. However you measure it, the answer will involve algos using larger minimum fills and significantly increased use of the “I would” feature. Less TWAP and VWAP, more implementation shortfall.

And increasingly, the answer will involve avoiding algos altogether and seeking block liquidity instead, especially for orders with the most alpha. Nothing befuddles a quant quite like a block trade. It’s not that block trades are immune from statistical inference; they’re not. Block trades are more visible. They stand out more when they hit the tape. There’s information in the price that clears the market for block size. But the analysis is harder to do. Blocks have less predictive value in part because they’re not as rational as trades generated by algorithm. Blocks are driven by judgment and intuition. They’re outliers. Quants stay away from outliers.

Systems and methods for completing block trades exclude smaller players. Some large orders may still show up at the small-order algo party, but the reverse can’t be true. When large orders finally decide to seek each other again directly in size, in their own exclusive venues, on the telephone or in the dark, the small-order quants and high-frequency traders won’t get past the bouncer. 

The presence of high-frequency traders has changed the nature of trading in small sizes and in many ways has undermined the reasons why institutions embraced algorithmic trading in the first place. As a result, more and more buy-side firms are likely to avoid algos altogether and seek block liquidity instead.

There is a flaw that underlies the premise of algorithmic trading of institutional orders.

When first introduced to institutional equity traders, the idea of taking large orders and splitting them into smaller pieces made a lot of sense. But that was many years ago now. Before the introduction of algos more than a decade ago, there was a time when small orders and large orders hardly interacted. Small orders were routed to the NYSE DOT system, or found their way to Nasdaq wholesalers. Large orders were put together upstairs on the telephone, or in the Instinet machine.

The impetus for change came in the late ’90s, when new sources of small order flow exploded onto the scene – the rise of Internet brokers, the proliferation of day-traders, and layoff business from options market-makers and program traders. Not to mention, the buy-side embrace of order management technology, the introduction of the FIX protocol, the Manning Rule, Limit Order Display, and the Order Handling Rules.

Among larger buy-side firms, the early adopters began using broker technology to slice their large orders into small pieces that were indistinguishable from the general flow of small orders. They found new sources of liquidity and lowered their transaction costs.

That was how it started: Large orders put on small-order costumes and masks and attended the small-order ball. Later adopters found there were fewer and fewer revelers back at the large-order mixer. Old school traders found it hard to resist slicing large orders with algorithms, not only to interact with the huge wave of small orders but also because that’s where the other large orders were. Eventually, they too put on small-order masks and forsook their market for the Algo Ball. And this trend continued to grow because it was efficient and lowered trading costs.

And it worked very well for a long while; but it doesn’t work so well anymore. It doesn’t work because the presence of large orders masquerading as small orders has attracted a new kind of small order to the ball. High-frequency orders simply weren’t there when the algo revolution started. But they’re there now, and they aren’t leaving. Their presence has changed the nature of trading in small sizes and in many ways has undermined the reasons why institutions embraced algorithmic trading in the first place. And those masks aren’t fooling anyone anymore.

Maybe it’s time for buy-side traders to take off the masks and go back to their own party.

Set aside the metaphor and let’s talk about what’s really going on in the markets. The term we use for “not fooling anyone anymore” is “information leakage.” The traditional kind of leakage occurred when your broker shared information about your order with people he shouldn’t. Leakage nowadays is different. Who you tell – or who your routing technology tells – is still important. But in the high-frequency age, the most pernicious cause of leakage is predictability. Algos have specific objectives that make them inherently predictable. Hence the problem.

To fight predictability, algorithms employ strategies that try not to be so obvious. They randomly generate displayed quantities, and use randomized time intervals between order replenishment. They randomize trigger sizes and randomize distribution among venues. The problem is these random values are all generated between parameters that are set by the objective of the algo. Over time these parameters can be inferred because the algos generate so much data. Large orders sliced into small pieces create many orders, many trades, many data points – and many valid samples.

This is not leakage caused by dark pool IOIs or conditional order technology. It doesn’t happen because a counterparty fades on you. It’s not based on who you trade with. It’s caused by your trading. Even an algo that avoids HFT counterparties still leaves a trail of data.

And your data gets analyzed, make no mistake. Take VPIN, which is highlighted in an article by Cornell professor Maureen O’Hara in the current issue of Financial Analysts Journal. VPIN, which stands for “volume synchronized probability of informed trading,” is a real-time measure that seeks to capture the probability that somewhere out there is an informed trader. Or a large order. You have to ask yourself: If this is what’s being published, can there be any doubt that even scarier stuff is already deployed against you in the marketplace?

Going forward, traders will minimize this forensic leakage by adopting strategies with broader parameters, less predictability and fewer trade executions printed to the tape. There are multiple ways to measure this, any of which might emerge as a standard. Average trade size doesn’t meet the requirement, as it obscures real differences among stocks and order sizes. But firms might choose to track the number of separate executions per completed order, or per hour, or per 100,000 shares traded. You could measure how many executions per million dollars traded, or per percentage point of ADV. However you measure it, the answer will involve algos using larger minimum fills and significantly increased use of the “I would” feature. Less TWAP and VWAP, more implementation shortfall.

And increasingly, the answer will involve avoiding algos altogether and seeking block liquidity instead, especially for orders with the most alpha. Nothing befuddles a quant quite like a block trade. It’s not that block trades are immune from statistical inference; they’re not. Block trades are more visible. They stand out more when they hit the tape. There’s information in the price that clears the market for block size. But the analysis is harder to do. Blocks have less predictive value in part because they’re not as rational as trades generated by algorithm. Blocks are driven by judgment and intuition. They’re outliers. Quants stay away from outliers.

Systems and methods for completing block trades exclude smaller players. Some large orders may still show up at the small-order algo party, but the reverse can’t be true. When large orders finally decide to seek each other again directly in size, in their own exclusive venues, on the telephone or in the dark, the small-order quants and high-frequency traders won’t get past the bouncer.