Predicting market impact is a prime use case of TCA. By having an idea of what the market impact of an order will be, traders and portfolio managers can make better decisions around how to implement their investment ideas. So how do we build a pre-trade model?
Let’s start by describing what exactly we mean by market impact. Market impact of an order is the influence of that order on the behaviour of the market. By behaviour, we usually mean the change in the ‘price’ of a stock and often use the mid-point of the bid/ask quote to represent the ‘price’. To measure market impact, we need to compare two things: (1) what happened to the market when we interacted it, and (2) what did the market do when we didn’t interact with it.
There is an obvious problem with this approach. We need to measure both (1) and (2) over the same period. But we can only do either (1) or (2), we can’t do both at the same time. So, what can we do? We will use a statistical approach. We will compare our predictions with our outcomes over many orders to see if the model does a good job in explaining the range of outcomes we get. Implicit in this approach is an assumption that the background ‘noise’ – i.e., the price changes that aren’t associated with our presence - averages to zero.
Our objective in building the impact model is to predict how the market for a given security will respond to how we interact with the market. We first note that each security has a unique liquidity profile. This profile changes over time. We can think of each security as having its own personality. The key aspects of the personality are things like spread, depth of order books, typical traded volume, and volatility. When we go to construct our pre-trade model, we want to tailor it to the personalities of each of the listed securities we have market data for. We are particularly interested in how easily the stock in question can get perturbed by our presence. Will it get volatile as we touch it with our orders, or will it stay quiescent? We seek to quantify this aspect of the stock at a security-specific granularity.
We take an approach of ‘passively’ looking at the market without any of our orders interacting. We can be sure that we won’t have any market impact when we aren’t interacting with the market (this is from our definition of market impact), allowing us to get a baseline for the way a particular security behaves when we aren’t trading.
We also know that as we are watching the market from the sidelines (i.e., when we aren’t trading), others most definitely are trading, and we can observe the responses of the market to events like incoming marketable orders. While we can’t directly see the incoming marketable orders directly, we can see the effect (impact) they have as they generate trades that are reported to the tape. We can also measure changes to the bid and ask sizes and prices, i.e., the changes to the order books. We can use this information to get an idea of how the stock in question responds to new orders. We need to be thinking about the ranges of outcomes here as our measurements of the responses produces a range of results. Hence the need for a statistical approach.
There is a very important point here that, when we make predictions we are not predicting specific numbers, instead we are predicting distributions. What this means is, we don't expect the impact to be a certain number, we expect the impact to have a range of possibilities. This reflects the statistical nature of our prediction - we predict both the expectation value and the confidence interval. The expectation value is the impact we would expect over a large number of orders while the confidence interval is a measure of the range of outcomes we would expect.
Realistically, for most stocks and for typical institutional sized orders, the confidence interval tends to dominate the expectation value - we are more comfortable estimating a range of outcomes rather than a specific outcome. The reason for this is while we know with some certainty how we expect to interact with the market, we don't know how other participants will behave. That said, we can, by watching the tape, have a good idea about the 'average' behaviour of other participants.
The next thing we need to understand is our intended trading strategy that we want to make a prediction for. Our model will need to take the intended trading strategy as an input. Things like how many shares to trade, how to vary the participation rate, etc… We need to know this info so we can use our baseline personality assessment of the security we are predicting to see how it would react to the way are intending to interact with it. Are we trading a large percentage aggressively or are we trading passively over a long time? Based on watching how the stock has traded in the recent past, we can make some educated guesses as to how the stock will react to our intended strategy. These educated guesses are what we call our model.
We can, of course, vary the order input to the impact prediction model allowing scenario analysis where we can see how our expected cost changes based on our intended strategy. Traders and PMs can get an idea of what the best way to optimize their alpha with the expect impact costs by varying the trading strategy and observing the change in expected impact costs.
Figure 1. We show scenario analysis with a pre-trade model to see the expected impact as we vary the strategy. On the left we show two strategies, a front-loaded and a rear-loaded strategy that shows the expected number of shares to execute over time. On the right we show the expected impact as the strategy is executed.
We also want to be able to adjust our post-trade performance. We can take the actual strategy that we executed and feed it into the model. We can also adjust the liquidity factors to reflect the experienced liquidity rather than the predicted liquidity we used in the pre-trade phase.
The final step is to check that our model is working. We can do this by taking what happened as seen by our post-trade TCA analysis and comparing it with what the model would have predicted. This procedure, called a z-score analysis, can tell us whether our model is doing a good job. If the model is not doing a good job, then we need to adjust the model.
In conclusion, when we build a model to predict the impact of an order, we use stock specific liquidity characteristics. We use a passive approach to watch how the stock interacts so we can measure not only the liquidity analytics like spread, depth and volatility but also how these metrics respond to trades (which can only come from marketable orders, but we can’t see these orders). By seeing how other orders affect the liquidity dynamics. We then take this knowledge and apply it to the way we intend to interact with the market. Finally, we check our results on a continuous basis by comparing the results we achieved (from our post-trade TCA) with our predictions and continuously re-tune the model based on our checks to ensure the model is as accurate as possible.
With a prediction of order impact in hand, we can improve portfolio optimization, run scenario analysis to determine the best implementation strategy, generate alerts when orders have more impact than expected and many more use cases. We can also demonstrate rigorous adherence to our best execution policies and procedures by trading best for our stakeholders.
Comments