A fresh look at pairs trading
Originally pioneered in the ‘80s, pairs trading has become a staple amongst a wide range of investors – from day traders to tier-one asset managers. Despite its ubiquity, the approach to executing pairs has changed very little over time. While algo providers have put huge effort into developing liquidity-seeking algorithms and improving benchmark performance; until recently, very little has changed for pairs.
To better understand the current state of pairs trading, I spoke with a range of practitioners across the buy-side and sell-side to hear their thoughts. In this paper we look at the difficulties of pairs trading and how a new approach can provide real benefits.
The Buy-Side Problem
One of the difficulties when deciding how to trade a pair is the wide variety of underlying strategies. Among the buy-side firms I spoke to, all had slightly different drivers behind their use of pairs.
For more passive funds there may be little-to-no correlation between the stocks – one side is sold simply to cover the cost of purchasing the other. At the other end of the spectrum (e.g. during mergers and acquisitions) the prices of the two sides will have a direct impact on each other. In the middle is the more traditional pair, two stocks with a strong historical correlation but no direct causal link between their prices.
Each of these requires a slightly different approach. For a basic cash-neutral pair, both sides may be equally liquid and able to be traded in a similar fashion. For others, one side will often be a lot more liquid than the other meaning that different strategies are required. For pairs of tightly coupled stocks, every trade on one side will directly impact the price of the other.
The end goal for the buy-side is the same, complete both sides of the pair, with minimal impact and without getting legged. But how can a broker achieve this when they don’t necessarily know what the client is trying to achieve?
The Sell-Side Solution
The traditional solution is to break the pair into two separate orders and trade each one with a simple POV strategy. If one side is more difficult to trade, then that one leads and, as it fills, the algo will catch up by trading the more liquid side of the pair. This has been the approach for years and while single-stock algorithms have evolved over time, the pairs algorithm seems to have been left behind.
The evolution of single-stock benchmark algorithms has involved better predictions of market volumes and access to a wider range of liquidity. Together these have led to better price performance. However, these same developments are difficult to apply to a pair.
Since pairs algos never know when the order may be pulled, or what might happen in the market, they need to keep both sides of the trade in-line at all times. The standard approach of waiting for the less-liquid side to fill means that, more often than not, it is forced to cross the spread to catch-up in the more liquid stock. Having a better prediction of the volume curve doesn’t help if the placement of your orders isn’t based on that curve.
Similarly, while new forms of liquidity are appearing, the majority has little-to-no pre-trade transparency. When a pairs algorithm is forced to keep two orders in-line, it is difficult to take advantage of the cheaper, but less-predictable liquidity that can be found in dark pools, periodic auctions or conditional venues.
For many brokers, these problems were too difficult to overcome and, even today, many pairs algos are restricted to just trading on the lit markets or performing a simple SOR sweep. Something that would be unthinkable for a single-stock strategy.
The Importance of Discretion
The difficulty in designing a performant pairs strategy is driven by two main factors. Firstly, there is the need to articulate the driver behind the trades; algos will typically support a couple of simple models or, worse, trade all pairs in the same way based on a simple ratio limit. This doesn’t reflect the level of complexity behind the decision to trade and the potential impact on the trading style.
Secondly, there is the risk of being legged. A buy-side trader needs to maintain oversight of their orders and have the ability to pull back or change strategy at any point. This means that the algo needs to keep both sides in-line throughout the trading day. Typically, algos are only allowed to get between 1% and 5% ahead on either side.
Adding tight volume limits to an algo forces it to keep catching-up, which generally means trading before it’s optimal to do so; ultimately hurting the order’s performance. Allowing the strategy to define its own pace can still get the order completed in the same time but gives more opportunity for price improvement.
Unfortunately, it all comes down to trust. Giving the algo enough discretion to get the best price means trusting that it makes the right decision. Since even the most advanced algorithms can’t process news and market events the way a human will, the risk of a bad decision is too high. No trader on either side of the fence wants an algo completing one side of a pair early, just in case.
Humans v Machines
The matter of trust is an important one. Especially as technologies such as machine-learning become more mainstream. Machine Learning has been around for many years and now that the hardware is available to support the millions of calculations required it is moving into everyday use.
Computers nowadays are very good at recognising patterns; whether that’s in speech, images or structured data such as stock prices. They can convert the words ‘Alexa, what’s the weather today?’ to a meaningful command, tell whether a photo contains a cat or a dog, and recognise a ‘crumbling quote’ across the US markets.
Getting to this level of pattern recognition isn’t easy. Machine learning algorithms require vast quantities of data for training and back-testing. You need to feed a computer millions of images of cats before it can start to recognise them on its own. While the financial industry does produce a huge amount of data it’s not always well suited. The constantly changing landscape means that a strategy trained on data from the last 5 years won’t necessarily be relevant right now. A cat is always a cat, but financial markets are completely overhauled every couple of years.
There is a big leap from pattern recognition to predictive analytics. Recognising a crumbling quote may tell you that there is a slightly higher chance of a price move, but it won’t tell you whether a stock will close higher or lower than it opens, or whether a news event is going to cause a jump in the price. Machine learning provides great benefits in terms of the venue-level interaction but is not quite ready to make the big decisions.
Despite these challenges, it’s reasonable to assume that technology will soon get to the point where it can provide a better long-term prediction than even an experienced trader. The technical difficulties are not the only challenge though, the harder one to overcome is the issue of trust.
Humans naturally give each other a second chance. If a trader makes the wrong call but had the right intentions, you might give them another order. If an algo does the same thing we are a lot less forgiving. A strategy that is right ‘most’ of the time will simply not be good enough.
A Different Approach
The solution to all of these problems is simple. You need to combine the intelligent liquidity access that can only be achieved with automated trading, with the insight of an experienced trader. You need machines and humans working together.
Following a more traditional ‘high-touch’ model means that there is always a human available to make the difficult calls. If a block becomes available, they can weigh up the benefit against the risk of completing the other half; they can consider the client’s strategy and preferred trading style in a way an algo can’t.
By combining this human oversight with modern liquidity capture tools, it’s possible to take advantage of all the new styles of liquidity without adding any undue legging risk. The overall result is a pair that’s traded quicker, at a better price and, hopefully, more in-line with the client’s underlying strategy.
The Analysts Challenge
Another reason that poor-performing pairs algorithms have been pervasive for so long is that there is no standard way of measuring performance. Pairs algos are traditionally marketed on their features rather than their performance, for the simple reason that no-one can say what a ‘good performance’ on a pair looks like.
Traditional measures like VWAP become surprisingly complex. For a standard limit order it’s simple enough to know whether a print was in or out of limit. For a pair, you need to look at the price of both stocks at the time of the trade and compare this to the target ratio if there is one. You also need to know whether the algo was allowed to participate in that stock based on the volume it had filled in the other. This ends up being such a restrictive approach that it delivers little value as a benchmark.
Other measures might be simpler but are naturally biased by the way pairs algos trade. Take a cash-neutral pair with one illiquid leg, and one liquid leg. The traditional approach is to trade the less liquid part passively and keep-up with aggressive trades in the more liquid stock. Since it is cash neutral, both sides have the same value and the same weighting, but the illiquid leg is likely to have a higher spread. Therefore, any average across the two will favour the passive side and potentially hide any impact from the aggressive trades.
On a large enough dataset, these may average out, but for individual trades they make it almost impossible to tell a well-executed pair form a poorly performing one. Instead of assuming that the same set of measures can apply to all types of strategy, it’s important that both the buy- and sell-side decide what metric they’re targeting and properly measure it.
The Perfect Pairing
From all my discussions, it has become clear how important pairs remain in modern trading. Whether the result of a closely guarded quantitative strategy, or a simple cash-neutral trade – pairs need to be traded very carefully. Despite this, the traditional pairs algo has lacked investment from most brokers.
Even those who have invested are still hamstrung by restrictive limits which prevent them from making the most of modern liquidity sources. The hesitancy to give discretion to an algo, and difficulty in proving performance even if you do, means it will still be years before an algo is able to trade a pair as effectively as a human trader.
The answer is to provide the best of both worlds; augment the insight and decision making of an experienced trader with modern tools and cutting-edge liquidity access. A buy-side trader doesn’t want to pick between a human and a machine, they just want the best outcome for their order.