The latest in LOXM and why we shouldn’t be using single stock algos

Algo trading, e-Trading and Machine Learning are becoming increasingly important topics in the quant world. Quant-based strategies have enjoyed somewhat of a resurgence in recent years, and that, coupled with exponential improvements in computing power, has led to some very interesting developments.

At QuantMinds International in Lisbon, an entire stream was dedicated to the subject, including talks on Bayesian asset pricing for algorithmic trading; algorithmic indices: how to build strategies matching the views of the client without any overdose fitting; and using AI for trade anomaly detection.

But perhaps one of the most hotly anticipated presentations was the one given by JP Morgan Chase & Co on the latest developments in LOXM.

Artificial intelligence using deep reinforcement learning

Heralded as the beginning of Wall Street's robot revolution, LOXM is the bank’s AI programme that executes client orders as fast as possible and at the best possible price, having been taught to do so from billions of past trades, both real and simulated.

And it has been very successful. Vaslav Glukhov, Head of EMEA e-Trading Quantitative Research at JPMorgan, told QuantMinds about the background to the project, and how they developed the system.

Glukhov explained that designing the programme was no mean feat.

The challenge is doing the best execution for clients while also keeping regulators happy.

Efficient order execution is a non-trivial problem and there are numerous reasons for that. For instance, the fact that different clients need different things and they all have to be served well. On top of that, market conditions are variable, not just in terms of price but also in terms of execution, and regulation is a heavy burden for banks.

“The challenge is doing the best execution for clients while also keeping regulators happy,” said Glukhov.

“You have to develop a product that does all these things, and that’s a daunting task.”

Limit order placement

In this environment, Glukhov said that they concentrated their AI agent in the limit order placement. “LOXM does very specific things, but it does them well,” he said.

“An AI agent must learn how to operate in the environment of bid/ask prices, and the liquidity availability on both sides of the book,” he explained.

“A buy order needs to be placed efficiently to not only take advantage of liquidity, but also to predict the future to the extent possible.

Glukhov explained how these were just some of the things that the AI agent had learned during training.

In developing LOXM, JP Morgan had employed standard reinforcement learning.

“In this approach we have an action, and the action is how much to place, what price, and for how long. It makes sense for the agent to be intelligent about the quantity that it asks for, and it needs to be smart in terms of how long it needs to place that order – if it gives the order for too long it will lose the opportunity and it will need to execute at a higher price. All these things need to be taken into consideration and the agent needs to be aware of the consequences of each action,” he explained.

LOXM was also designed with specific objectives in mind.

“We designed it with specific business objectives in mind, specific market conditions, specific sizes and tickers, and external constraints, of which there are numerous.

“How LOXM is rewarded for being efficient in the market, and how the efficiency of the agency is defined, is stated in the reward function,” he explained.

“We send simulated orders to the exchange, we simulate how they execute, we simulate market impact, and then we feed the reward and batches of execution back to the agent’s brain, which memorises which actions are good and which are not so good.

“On the version we are running, we use the standard RL approach which is based on dynamic programming and Bellman equation.

“The total reward in the agent operation is not necessarily the sum of local rewards,” he added.

Lessons learned and questions remaining

Among the lessons Glukhov said they had learned was the need to be very specific in how they train the agent.

“In traditional quant finance, we try to be as general as possible. We are trying to develop theories that resemble physics in some respect. This works great to an extent, but we lose the idea that markets are very specific, they are very competitive and they are Darwinian. With AI, we can do better, we can create things that handle very specific situations very well.”

Glukhov explained that another challenge in building the system was that they had to be able to explain what the agent is doing.

“We want to be able to explain it to our clients as well as to regulators and prove that the algos won’t disrupt the markets. We can’t just tell them we did that because we trained the agent and that’s what the machine tells us.”

Glukhov ended his presentation with a note of caution:

“Taking all this into account, it is not self-evident that the perspective of deep learning, in electronic trading at least, is very clear. We don’t know that yet, that’s why we are moving very cautiously, moving step by step, exploring things, and not trying to do all things at once.”

The fallacy of using single stock algos for portfolio trading

The afternoon was not only dominated by machine learning. We also heard from Michael Steliaros, Global Head of Quantitative Execution Services at Goldman Sachs, about the fallacy of using single stock algos for portfolio trading.

Passive strategies are now 35% of the volume traded in the US market. That figure was 5% 10 years ago.

Since moving to the sell-side (Steliaros spent more than a decade on the buy-side building quant stock-selection models and managing global market neutral equity portfolios), Steliaros bemoaned the fact that he had to use single stock algos.

“Nobody could execute with correlation in mind.”

“Even when we did have products like that from brokers, they were typically based on the same Bara-type models we used in portfolio construction, with long horizons, averages across volatility correlation, and market impact estimates that were not really suitable for intraday trading.”

New models were needed, he explained, because of the change in intra-day trading dynamics brought about by the explosion of passive, ETFs and smart beta products.

Intra-day trading dynamics

For instance, there had been an increase in the percentage of client flow driven by quant overlays, and a very large percentage of assets being traded day in and day out across markets globally were quant driven. “Passive strategies are now 35% of the volume traded in the US market. That figure was 5% 10 years ago,” he stated.

What they all had in common, he added, is the fact that they have a market and close benchmark. This, he said, had shifted the intra-day volume profile significantly, as well as having shifted the dynamics of trading over the course of a day or multiple days equally significantly.

In addition, a portfolio approach needs execution to be aligned.

“If you trade with a single stocks algorithm you will trade your very liquid stocks faster than your less liquid ones, and the original trading list that may have been sector neutral ends the day tilted, so you are exposed to risks that cannot be managed.”

The other important thing to consider was the rise in transaction costs. “Cost was a constant we put in our models against the alpha, but today that’s no longer the case.”

Building blocks

Among the things we needed to think about when trading lists of stocks, was to formulate an adaptive portfolio trade scheduling.

Another thing to consider was that clients could well have multiple benchmarks, and that had to be taken into account.

Risk had to have an intraday view, and there was also the portfolio level risk to take into account.

Market impact was another input in the decision making of how to trade, argued Steliaros.

“The market impact is typically estimated on daily averages. What one needs to do given these changes in intra-day dynamics is monitor limited order book dynamics, see who the participants are that you are acting against slice by slice, and given the fact that there are days when high frequency market makers could be 70-80% of the market, that’s a very different type of liquidity you are accessing. We need to be able to model how we access and where the liquidity that we’re accessing comes from.”

As the chairman of the afternoon stream so eloquently reminded us: “We are here because we are passionate about maths, technology and finance. We are also passionate about learning. It’s exciting to be a quant we always have new boundaries to explore.”