BMLL partners with quants for HFT regulation

BMLL Technologies has partnered with a team of quantitative researchers from the Paris-based Ecole Polytechnique, providing the academics with order book data for their research, which takes a statistical approach to markets regulation.

Mathieu Rosenbaum is a senior professor at the school, where he holds the Analytics and Models for Regulation chair. His research focuses on statistical finance, mainly modeling market microstructure, and looking at ways to help financial regulators adapt to high-tech markets.

“We are trying to regulate very technological, very clever, very advanced, very scientific market participants. So we believe that we should also adopt a financial engineering approach to regulation,” Rosenbaum says of his research.

BMLL is providing the quants with “Level 3” data, the use of its Data Lab platform, and access to its data scientists. Data Lab is a cloud-based Python environment where users can access the data, as well as analytics libraries. Rosenbaum and his team have been working with BMLL data since Q1 of this year.

“Level 3” data is what BMLL calls data that goes beyond the top of the order book. “It’s every single message coming out of a public exchange. It’s not just seeing what the best price is, what a trade is; it’s every single insertion, deletion and cancellation message that comes through an exchange on any given day. That is a vast amount of data,” says the company’s chief product officer, Elliot Banks. Banks says that BMLL’s dataset weighs in at a hefty 15 petabytes. For scale, Netflix’s data warehouse is 60 petabytes.

BMLL takes in the data as a disparate dataset from exchanges, and puts it into different products, one of which is Data Lab. Data Lab is mainly aimed at market participants who want to use it for best execution, to understand the market as an exchange participant, or for alpha generation.

The Ecole Polytechnique quants, however, are looking to use the data in models they build to understand market microstructure—the complex interactions of price discovery, trading behavior, and trading venue structure—in a high-frequency trading (HFT) context. A microstructural approach looks at all the mechanisms that play a role in price formation, the way the market functions at its core, plus all the actors and events that influence this process.

“As academics and researchers we can find a way to look at data and say, OK what is the impact of making that choice, if I choose to have a given tick [the minimum variation of prices]—1c or 5c, for example—how will that change the markets? How can I decide when to interrupt continuous trading and trigger an auction?” says Marcos Carreira, a PhD candidate in Rosenbaum’s program and former technical modelling officer at the Brazilian exchange B3. “So we have good models that tell us what the behavior of investors is supposed to be, and then we can go and get data to see if the model fits reality. And then we can take our conclusions to the regulators so they can look at the different possibilities of organizing the markets.”

More traditional approaches to markets might ignore these events, which happen so fast they are measured in microseconds and milliseconds, such tiny slivers of time that only computers can understand them. Many academics in the past, for example, have been largely focused on valuation, looking at the fair market prices of securities, and projecting returns. But these approaches cannot explain how the market might change instantly with new information and new customer orders.  

“If you were interested in long-term problems, you might neglect or not really care about them, but in fact what is happening at the microstructure level has an impact on everything. Even long-term volatility is connected to what is happening at that level,” Rosenbaum says.  

In the HFT world, market makers’ behavior is highly dependent on volatility, a topical area of study in the past year, which saw unprecedented volatility. Rosenbaum himself, and along with a colleague, developed a groundbreaking model for “rough volatility,” solving a long-standing puzzle in options markets. But he and his team are also working with regulators on other projects, which is where they hope the BMLL data will make a difference.

“We need to understand some very specific events—Brexit, the Flash Crash—and it would not make sense for us to do some quick and dirty statistical study. You need to take the time to have full information from the data, and this is where BMLL data comes into play,” Rosenbaum says.

Regulators going quant

The regulatory and supervisory world is a traditional, even hidebound place. However, regulators are starting to use emerging technologies to become more efficient at keeping markets safe and fair.

“In the past, the only academics that regulators were talking to were well-established economists. But that is changing. In the case of extreme volatility, for example, the job of the regulator is to understand what happened, and how they can mitigate the effects of it. Well, you need more than a simple economics model to do that, you have to dig into the data and use complex mathematical models, you have to get people able to do that, and this is where we could collaborate with them,” Rosenbaum says.

Rosenbaum’s team collaborates with regulators such as the French markets regulator, the Autorite des marches financiers (AMF). Late last year, Rosenbaum and others published a paper co-authored with AMF managing director Philippe Guillot in which they laid out a new matching design for financial transactions in an electronic market. The mechanism, which they dubbed Ahead (which stands for “ad hoc electronic auction design”), allows market participants to trade between themselves at a fixed price and trigger an auction when they are no longer satisfied with this fixed price.

The researchers said in the paper that they proved that Ahead worked better than the central limit order book (Clob), the standard approach that many exchanges use. Since HFT became more widespread, there has been a debate that Clob is a suitable matching mechanism.

While Rosenbaum could not divulge the details of projects that he and his team are working on with the regulators using BMLL data, he says they’re developing similar kinds of research questions to take to the regulators. With the Level 3 order book data, for example, the Ecole Polytechnique quants could calibrate an order book model to assess and validate the toxicity of order flow. The Level 3 data is necessary because “if you want to investigate HFT strategies, you have to look at flows at least at Level 3,” Rosenbaum says.

Ever since HFT has come into the public consciousness via events like the Flash Crash, there has been a lot of debate in regulatory circles about whether or not it is good or bad for markets. But, Rosenbaum says, HFT is such a wide field, such a multiplicity of strategies, that to tar it all with the same brush is not helpful to regulators. A financial engineering approach could distinguish between trading firms’ individual strategies. 

“Proprietary trading firms, HF market-makers, and even asset managers are not just putting order books at the bid or the best ask, or putting in market orders. Their strategies are complex, and you need to consider at least two or three levels down into the order book. So we use data to classify the order flow, and whether this order flow is stabilizing the market, or, on the contrary, this order flow is destabilizing the market,” Rosenbaum says. “And from a regulatory perspective that is very interesting, because it’s not just saying, OK you are putting in a lot of orders so you are good, or you are bad.”

Carreira says that besides the Level 3 data, BMLL’s Data Lab was attractive to the research team partly because it functions in a centralized workspace. If he were to leave the program, for example, the rest of the team could still access his code. It wouldn’t be lost in a folder or on a device somewhere. More critically, it also gives centralized access to the dataset, meaning that future researchers can repeat his work, allow them to test and reproduce his results.

He adds that he feels the BMLL data is of high quality. The data the team uses must be able to accurately present a picture of events. A researcher, for example, might want to do a relatively simple study, and match up the price of a transaction that occurred at a particular time and the bid/ask at that moment of execution. Two exchanges might record that data very differently, however.

“I’ve worked with market data directly from exchanges, but the ability to work in a well-structured Python environment like BMLL, with data correctly classified and labelled and the ability to summon snapshots and events on the order book with just a few lines of code, means that my time is spent on features and dynamics instead of reconciling timestamps and identifiers,” Carreira says.