An Invitation To Pool Data

Within equity trading at APG, we tend to think of parent orders as falling within three categories. The first category is orders which can be executed within one business day, usually under 6% of the average daily volume (ADV). The second category is orders which can complete within one week. The ADVs of these orders usually range between 10% and 45%. The third category of orders spans multiple weeks and their original ADVs can be multiples of a day’s typical volume.

Under-one-day, under-one-week, and multiple weeks. For each of these categories, the goal is the same: to get the best price possible for the portfolio manager (PM). Within the “under-one-day” category, the process tends to focus on the algorithm providers - wheels, algorithm suite choice, algorithm choice, algorithm customization, VP and price caps, auction usage, etc. For the “multiple-weeks” category, we have been reaching out to other trading desks to discuss an alternate workflow that we think has the potential to lower trading costs, so please feel free to contact us about that.

Under-one-week

When we think about the “under-one-week” category, we start to think about the problem in the following way: suppose the share price has moved 40bp away from where it was trading when we began the order two days ago, and we’ve been 9% of volume thus far, most of which was executed in the dark, and our residual ADV is 15% - what is the best thing to do at this time? Should we be seeking blocks, should we be out of the market, or something in between?

Note in the above there isn’t any mention of the particular PM or strategy. There are cases where the PM would prefer a quicker execution schedule, but these are exceptions. In most cases, one can assume the alpha the PM is targeting is likely to take months to unfold and when it’s otherwise it can’t usually be identified ahead of time. There are several reasons for this which we won’t go into here.

More formally, we would define the situation the trader is in at any point in time by three features of the parent order: the residual ADV, the change in the price of the security, and the volume participation rate (VP). Note that each of these is relative to when the parent order began so the VP rate would be all executions divided by all market volume since the arrival date and time, and the change in the price of the security is vs the same arrival date and time. In regards to the actions the trader can take, we think of them as VP bins: the trader could seek liquidity, the trader could do nothing, and the trader could be between 4 and 8% of the volume or 8-15%.

In the language of reinforcement learning (RL), we think the state space has these three dimensions, and the actions collapse to just four. Once the analysis has been done of what state the trader is in, there should be an optimal policy that says whether the trader should seek liquidity, do nothing, or something in between. The chosen action maps to an algorithm wheel with price and VP caps.

The only thing required for RL algorithms is a clear goal and lots of parent order data in a time-series format to train on. The goal we’ve been using - likely the same as yours - is to get the best price possible over the one-week horizon, but also take into account what the share price did in the two-to-five days after we finished.

Context

Some will ask: why do all this? At APG over 80% of the notional value we execute in equities is via broker-provider algorithms, and like other large asset managers, the parent order ADVs being traded need to be sliced into multiple algorithmic “placements” over a multi-day window. Ideally, the goals of those placements would be the same as that of the PM who generated the parent order. In reality, the goals are correlated, but not the same. The reason is that brokers are not optimizing to multi-day performance, rather they are optimizing to algorithm performance over the period of time they advertise their algorithm is likely to use to complete.

When we ask brokers why they can’t deliver algorithms that are optimized to our multi-day goals the answer we get is they lack the data. They lack the data to train these types of algorithms because clients only trade a portion of their parent orders with them, and clients won’t share the balance of their data, so the brokers - despite their years of experience - ostensibly aren’t building up the same experience that buy-side traders are. Buy-side traders learn a multi-day experience of interaction with a security price and brokers develop a very telescoped approach to execution, divorced from the context of the multi-day order.

What this means is the execution algorithm has the same logic on day one as day four of our parent order. Its behavior might differ, but not by much. It has no sense of the history of the parent order. Take auction usage for instance. The algorithm will typically ask the buy-side trader how many shares he or she wants to execute in the auction, or to use what price limit when approaching the auction, or maybe the algorithm is smart enough to grow its participation based on where the security was trading before the auction. But for buy-side traders with a multi-day horizon, the question is: is it better to push the auction price now in exchange for more volume or wait until the next morning of trading? The algorithm providers can’t answer the question definitively without access to everything the client did that day, in the auction, and the next day’s first hour of trading. That would be the minimum data set necessary to be analyzed across thousands of parent orders.

Private Information

When Liquidnet launched in the early 2000s it was initially met with skepticism by many buy-side traders. Shortly after, there was a golden age of a few years where traders would stand still and trade much of their parent order at once. After that, the behavior slowly deteriorated. It deteriorated because what was rational for individual traders was not rational for the larger community. Suppose a buy-side trader had been waiting in dark pools and crossing networks for two days with a certain minimum fill size, for - say - 15% of his or her entire parent order. After two days of no fills suddenly the entire placement is executed. What information does that convey? It conveys there’s a new participant in the name. The rational thing to do is not to reload the 85% residual but rather to wait. Let the new participant push the price a little and then circle back to them and try to resume crossing. No one wants to get run over, but everyone is worse off because block liquidity is less. Fast forward to today and everyone is crossing small order sizes, much to the consternation of PMs.

The point of the above is that there is contextual information the buy-side trader possesses, which isn’t publicly-available information like the level of VIX, the correlation between the SP500 and the oil price, or the average share movement on an earnings beat, etc. This private information pertains to the history of the parent order - the “picture” that the buy-side trader is aware of. It’s information like “typically when I’m 15% of volume for 20 minutes the share price goes against me, but in this case, I’m getting all the liquidity I want.” It’s the interplay between the actions the trader is taking and the response of the security. All of this data is being captured by the EMS and the broker whose algorithm is being used at that time. This data might be used by the algorithm to make decisions, but the algorithm doesn’t have access to the time-series history of the data since the beginning of the parent order to contextualize its decision-making.

Reinforcement Learning

So, that’s how we come to the features of the RL model we described at the beginning of this article: the participation rate and movement in the security neatly encapsulate that private information at any point in time; the VP bins are the actions that can be taken, and the residual ADV is what still needs to be done. We’ve had some success training and implementing these models, and we’d like to improve them with data spun off by all of our placements: unfilled dark orders, atypical venue liquidity, spread capture, time-of-day effects, etc.

We think it’s more efficient for everyone to have a tighter alignment of the goals of buy-side traders and their algorithmic providers. Algorithm providers should be encouraged to deliver solutions that are optimized to a multi-day window because that’s the period of discretion typically granted or needed by a PM. The data spun off by this multi-day interaction should be mined for performance, and this can be done on the buy-side trading desk, shared with the algorithmic provider, or a combination of both.

Benefits

As part of a larger “digitization” drive at APG, trading is taking a more data-driven approach. We have signed non-disclosure agreements with algorithmic providers to share data, we have onboarded providers who can interact with us in real-time via APIs, and we are in the process of changing our OMS and EMS to more API-centered providers. We have increased our expertise in quantitative and programming skills, and we have reached out to several parties to help us better analyze our data.

RL and deep learning algorithms need a lot of data, and the multi-year data we have may not be enough to arrive at a policy that approximates “ground truth.” And we suspect almost all buy-side trading desks are in the same situation. We each have experienced traders, PhD-level quant skills, our own limited data set, and algorithmic providers who are slightly misaligned to our needs. Our perspective is, that as an eco-system of data providers, analysts, buy-side traders, and algorithmic providers, we can move forward by some of us pooling our data to a common time-series standard of multi-day parent orders. We’ve already seen benefits from doing this internally. Pooled datasets yield insights and benefits such as:

• Lower slippage, hence better portfolio performance
• Genuine algorithm and venue evaluation
• Automation of a large part of under-one-week parent orders
• A better understanding of how and when to efficiently move large ADV orders through the market
• Publish an academic paper, get more feedback
• Encourage the algorithmic suppliers to up their game - match their products with buy-side needs
• Leverage the data in the EMS as it was meant to be used
• Make our behavior in the market less discernible to fast-moving players in the market who seek to profit from larger ADV orders

Data Anonymity

The format of any pooled time-series data would look something like the following.

In the above table, there is no ticker, direction, or account information, and five of the six features are either private or semi-private information to the multi-day parent order.

If we pool this kind of data across participants and use it to train models, this raises some natural concerns around data privacy. These can be addressed by a variety of approaches. There are legal tools like non-disclosure agreements, as well as technical tools ranging from the simple to the highly complex. On the simple end, removing or masking any fields that aren’t necessary and other forms of pre-processing and normalization can help reduce the information content that is pooled to what is most essential for the modeling process. Beyond this, there are also techniques for data perturbation or synthetic data generation that can preserve much of the usefulness of the data for a certain kind of modeling, while greatly reducing the potential usefulness of the data for other unintended purposes. Overall, the level of data privacy provided is a spectrum, and somewhere on that spectrum is a reasonable point that will provide our ecosystem with much more value than the level of risk it represents.

Conclusion

We share a common goal of improving trading performance for the benefit of our end clients. Our perspective is to break order flow into three separate categories with a different improvement approach for each: under-one-day, under-one-week, and multiple-weeks. With respect to the second category - orders which can typically be completed within 5-6 business days - we think a reinforcement learning paradigm is the right way to approach these orders. We’ve had some success with this approach and we think we and other asset managers can benefit more by pooling data to develop a rational approach to multi-day scheduling.

Based on analyses we've seen, there has been hardly any improvement in the implicit cost of trading large ADV orders in the last fifteen years. We shouldn’t be conservative, we should strive to make our ecosystem better for our end clients. Please reach out to us if you’d be interested in joining our research. And please ask about our alternate workflow approach to multi-week, high ADV orders.

Thank you for your time and attention.

And thank you to Allison Bishop at Proof Trading for her contributions to the above.