These robo-traders use quantitative models to decide how to execute orders in the market, often in a number of smaller pieces, thus allowing human traders to focus on the difficult, subjective side of the business that a machine will never be able to handle.

Algorithmic trading tends to focus on objectively measurable benchmarks; otherwise what different strategies aim to achieve becomes vague. There are five core strategies that aim to beat their respective benchmarks:

• volume-weighted average price (VWAP);

• time-weighted average price (TWAP);

•  market-on-close (MOC);

• implementation shortfall (IS) – the benchmark is either the mid-price when the order arrives or the opening price if the order is submitted outside market hours;

• participation – user-specified percentage of the total volume traded.

As in portfolio management, benchmarks often get a rather bad press but, even if the trader’s performance is not explicitly evaluated against them, they serve to describe the trading style of each individual algorithm. A VWAP strategy schedules a trade according to a historical average volume profile: it will break up a trade into smaller pieces, trading them at, say, every five minutes according to the volume that has traded historically in that interval. It is the natural way to trade over a given time interval to minimise impact. Participation is a more adaptive strategy: it controls market impact – how much you move the price – by trading according to the amount of volume executing in the market. Implementation shortfall can be summarised as “please work out how quickly to do my order to optimise the trade-off between price impact and exposure to adverse price movements or opportunity cost”. Many brokers now offer further parameters to allow traders who are not explicitly benchmarked to customise an order. Parameters are generally either dependent on price or volume, such as a maximum percentage of the volume traded or a limit that stops a stock from trading if it outperforms the index by, say, 2 per cent against the previous day’s close. These parameters can even be used to hybridise an order. VWAP with a maximum volume parameter of 30 per cent will have characteristics of both VWAP and participation orders – it will trade the lesser of VWAP or 30 per cent of the volume traded in the interval.
Vwap strategies

Figures 1 and 2 illustrate the use of different strategies and enthusiastic take-up of customisation for VWAP strategies. By strategy, VWAP accounts for 59 per cent of Citigroup’s flow, with one third of that constrained by either price or volume parameters. Vanilla VWAP orders – those with no additional parameters set – account for only 40 per cent of the total notional amount traded on Citigroup’s algorithms. As traders become more comfortable using algorithms, the fraction of the flow that is unconstrained vanilla VWAP is expected to continue to decrease.

FIGURE 1: Breakdown by benchmark of Citigroup’s algorithmic trading flow in 2005, showing VWAP as the most popular trading style

FIGURE 2: Breakdown of VWAP flow according to ‘vanilla flow’ (no constraints), and flow with additional price- and volume-dependent parameters

                                                            
                                                                     Source: Citigroup

Note: Figures 1 & 2 combined show that vanilla VWAP accounts for only 40% of Citigroup’s algorithmic flow

A word of caution: the phenomenon of ‘creeping featuritis’ is well-known – technology applications accrue additional buttons, switches and scrollbars without any apparent purpose. Any additional parameters should be useful rather than gimmicky, quantifiable rather than abstract, and never make the entry of simpler orders any more complex. Ideally, the interface between the user and the algorithm should be customisable so that parameters that will not be varied can either be taken away or set to a suitable default value. The trade-off between interface complexity and flexibility of the algorithm is essentially yet another efficient frontier: the interface should be optimised to be as simple as possible, given the degree of flexibility that the user needs.

Ideally, brokers should be able to offer pre-canned combinations of strategies and parameters to clients, so their favourite strategies can be accessed with a single click of the mouse.

Building blocks

Much is made of the complexity of designing algorithms: the terabytes of data that need to be sifted through, the need for holding a PhD in applied maths and so on. Yet the basic building blocks of any algorithm are no different to those a human being considers while trading an order: time, price and volume. The interplay of these three quantities varies hugely between markets, and even between different stocks within a particular market.

The brokers with a significant competitive advantage are those that have invested in bringing together regional teams to develop algorithms, rather than centralising their development globally.

For instance, a major difference between world markets, and even within Europe, is the importance of the bid-ask spread (the difference between highest-priced buy order and the lowest-priced sell order). Markets such as Italy, where the minimum price increment is large relative to the stock price – i.e. the minimum bid-ask spread is in tens of basis points (bps) (1bp is 0.01%) – behave radically different to the US market where the typical bid-ask spread is a few bps. Thus an algorithm optimised for the Italian market will perform badly in the US and vice-versa.

The Italian market is a good illustration of the usefulness of algorithms: the best bid and ask of a stock will often not move for many hours. The optimal strategy is to be patient, waiting on the order book. Manual trading of this requires constant watching in case the price starts moving away, leaving orders buried in the order book. A computer is able to watch the order book constantly at multiple levels and react instantaneously.

Algorithms will never replace human beings, either on the buy-side or the sell-side. The trading environment will change and will continue to do so as the use of algorithms becomes more widespread and dealers become more comfortable with them. The traders that adapt most successfully to these changes will be those who are able to make the most of situations where algorithms have an edge, while still using their own nous and market knowledge to deal with orders where needed.

Computers and humans

What advantages are there in using computers? Computers can monitor and react instantaneously to hundreds of market data updates per second, without needing to go for lunch or suffering from a hangover. Algorithms will always struggle to react to specific items of news, massive market shocks or other events that are impossible to model statistically – that is where the human being will always have a competitive advantage. This explains why algorithms are often touted as suitable for low-touch orders – those that can be stuck in the machine and left to run – and this is probably their main use at the moment. Their ability to deal with opportunistic trades is less well reported. Parameters, even such as a simple limit price, can be applied to most brokers’ solutions and so an order can be placed away from the market and left to trade – or not – without the need for the dealer to be transfixed by the market data ticker of that particular stock.

Algorithmic trading is meant to make life easier by freeing up time to work orders that require human intervention. To achieve that, it needs to be easy to use quickly and efficiently. Therefore, it should be integrated with the order management system (OMS) to avoid typing orders in twice, which is not only a hassle but also increases the probability of error. Similarly, trading objectives should be readily translated into the right strategy and parameter set, beyond simple benchmarking. Human language is incredibly powerful: the complexity and nuances of an order delivered over the telephone between human beings in just a few words can be nigh on impossible to encapsulate in a few numerical parameters. So algorithm designers need to devote just as much time and resources to modelling the market microstructure well as they do to systematising trading objectives and styles: brokers need to be prepared to work with the buy-side in partnership to customise their algorithms to their clients’ needs, either through new parameters or canning a combination of existing options.

Transparency and clarity are also vital. The algorithm should behave as expected, and the broker should be able to discuss how the model works at whatever level of detail is needed. Any model that is so convoluted that the broker cannot or will not explain its fundamental principles is unlikely to deliver what is expected or required.

Allied to transparency is the question of anonymity: who can see your flow? A key selling point of algorithms is that they can operate separately from brokers’ other trading operations, potentially with no human intervention, and thus ensure that information leakage is minimised.

The final criterion is performance: both the average performance of the algorithm and the execution volatility should be considered when choosing an algorithm. Execution performance is no different from portfolio management in that any improved performance goes hand-in-hand with increased risk. Choice of algorithm will therefore depend on how willing the trader is to be exposed to unsatisfactory fills on individual trades to perform better on average. Obviously, the more often a trader trades the closer their overall performance will be to the average performance of the algorithm. In the same way that the algorithmic trading providers should be integrated with the trader’s OMS, so should the latter’s pre-trade and post-trade applications, to maximise ease and efficiency of analysis and uploading trade data.

An application like Citigroup’s BECS (www.becsonline.com) is an ideal solution: historical executions from a large universe of market participants is collated and analysed, enabling easy performance comparison, and compared with forecasts made by a proprietary transaction cost model.

Algorithms have been proven to enhance both productivity and performance on the sell-side. The buy-side can now access these directly but is faced with many offerings and algorithms from which to choose. The best solutions are easy to use, transparent, flexible without being overly complex, anonymous, designed regionally, behave as expected and fulfil trading objectives, and they perform well against their peers. Once these criteria have been applied, there will no longer be such a wide choice.

Tom Middleton is head of the European Algorithmic Trading Team at Citigroup, of which Atchen Nathan is a member