ITG Adopts Machine Learning Techniques to Boost Fills in Dark Pools
Institutions that source liquidity in dark pools are often wary of interacting with small orders, since they are perceived to be information-seeking and potentially toxic. Rather than avoiding small orders or specific dark pools, ITG announced that it developed a new quantitative method for solving the problem of increasing fill rates in dark pools while filtering out potentially toxic liquidity.
On October 15, ITG went live with POSIT Marketplace 3.0, an algorithm built on a low-latency trading platform that uses optimization to allocate different order type combinations to 25 dark execution venues. "It generates a lot of orders and it generates a lot of complexity, so we needed a new technology approach to that," says Ben Polidore, managing director, Algorithmic Trading at ITG. The algorithm employs machine learning techniques and ITG's proprietary liquidity filter to source high-quality flow from nearly any dark pool, while minimizing information leakage, states the release.
"The goal of the product is to find every bit of natural liquidity from other investors and zero interaction with market makers and high frequency firms," says Polidore. At a recent equity conference, industry experts said the average order size in alternative trading systems is 187 shares.
Although 100 shares orders are often associated with HFT, Polidore notes, this is not always the case. It becomes difficult to tell orders apart, since investors use algorithms and these look very similar in some ways to high frequency players, said Polidore. "If you are using a VWAP algo, very often it is trading 100 shares at a time in a dark pool."
Working on the R&D for about 18 months, ITG drew upon its experience with optimization, in which it has a portfolio optimizer for clients and portfolio algorithms. "We used some of that knowledge to think about routing, instead of which stocks to trade, which order types to use."
While other brokers and third party order routing firms offer dark aggregation, ITG refers to the new method as "dark optimization." It built the optimization "as a generic function" to figure out how to split up the order across 25 dark venues including ITG's venue POSIT Alert. "If you are trying to route to 25 dark pools, it's not a tricky thing. You can split of the orders in a number of way," he said. When ITG added the order types, it needed a new technology and approach to manage the complexity, he said.
A tricky problem to solve
Since ITG uses four or five order types in each of the 25 dark pools, that quickly adds up to 100 order type/dark pool combinations. For example, the algo may send four or five order types to the same venue if this allows it to get at the counterparties that it thinks its customers want. "It's not 100 order types, but 100 order type/dark pool combinations," explains Polidore.
"It's much different than having 100 different order types, and saying let me find the way to allocate that. If you were given 100 pieces of papers, and told to put them into the tightest circle, it's tricky and it's very hard to come up with a mathematical function for that," said Polidore.
Polidore's team built stochastic optimization, a mathematical and statistical process that involves probability, to figure out where to send the order. "The optimizer figures out how to split up the order, "and when you look at the results it's very intuitive," said Polidore. His team doesn't need to tell the algorithm what to do. "We [tell] the optimization process -- the utility function -- here's the diminishing return of sending more and more stock to a particular order type and dark pool combination, and then it figures out the optimal number of places to go and figures out the order type,"explains Polidore. Clients can provide ITG with minimum order size constraints but ITG would like to have some discretion over the process, he notes.
Machine learning techniques
The algorithm learns while it trades based on historical data and decides what it will do now. After a period of time, it will realize that it didn't get any fills from a certain pool but that it did get fills from another pool. Then it figures out, "let me update the utility functions and re-optimize to come up with a different allocation," he explains.
"It figures out the optimal number of places to go and figures out the order type," he adds. "We basically try things and as we experience the response to those actions, it will change our behavior to the next iteration and it will continue improving as we have experience with the market."
Unlike a standard optimization approach where it looks at history and acts in a very predictable linear fashion, with machine learning, the algo designer needs to define fitness. "You try a bunch of random outputs and decide if they are fitter than the last iteration and then you sort out the fittest and have survival," he says.
Now that ITG has launched POSIT Marketplace 3.0 for U.S. equities, it's looking at applying the optimization approach to other markets with fragmentation across dark pools. "This approach definitely makes sense in Europe the way their markets have evolved. We just have to build it," said Polidore.
Ivy is Editor-at-Large for Advanced Trading and Wall Street & Technology. Ivy is responsible for writing in-depth feature articles, daily blogs and news articles with a focus on automated trading in the capital markets. As an industry expert, Ivy has reported on a myriad ... View Full Bio
Tweet This