With the expansion of algorithmic trading to the buyside, there are a number of potential pitfalls that an algorithm can fall into. It is important that users be kept aware of several problems - real or perceived - which can negatively affect the performance of an algorithmic trading strategy.
All algorithms require a market data feed – and here data latency issues may interfere with the design of the algorithm. This is primarily a technological problem, and it is addressed by getting better connections and direct data feeds, but at some point there is a noticeable trade-off between delivery speed and data quality. The fastest data feeds are the “dirtiest”, and occasional instances of faulty data being fed into an algorithmic strategy engine may affect its performance in an unpredictable way.
This problem necessitates data filtering procedures, to remove bad data from the data stream, and, as before speed is of the essence. A good filter should be smart enough to isolate the obvious bad pieces of data, but at the same time it cannot allow itself to become encumbered, since this will slow it down.
The quality of the historic market data is important for an algorithmic trader. It is used in the process of developing, testing and calibrating the strategy; it is essential for pre-trade analytics (offered now by many algorithmic brokers), and it is also used runtime to supply the analytic inputs to the strategy. Some of the data used in the strategies, such as stock volatilities or the correlation matrices, must be estimated using statistical methods.
This requires both creating and maintaining an extensive historic database of tick-by-tick data, and developing (or purchasing) statistical tools necessary to extract information from this database.
Data sources
In particular, all existing pre-trade tools require numerous inputs from a historic database – to calculate the average daily volume and the intraday volume distribution, and estimate expected stock volatility and bid-ask spreads. The accuracy of these estimates depends both on the quality of the historic database, from which the data is extracted, and the methods used to perform the calculation.
Although two pre-trade engines may use the same mathematical framework (which is more or less the case now), they can produce significantly different outputs based on different data quality and estimation techniques. This is something to keep in mind when comparing the predicted outcomes of different trading scenarios.
It is important to test all statistical data which is used in trading strategies. At Miletus, we validate both our data sampling procedures and the results of the calculations.
In light of the above, one is forced to recall a famous maxim by Einstein -- Everything should be made as simple as possible, but not simpler. Actually, both parts of this phrase are highly relevant to algorithmic trading. It does not pay to have an overly complicated strategy - it will almost certainly be highly sensitive to even minute data quality problems, and is unlikely to show stable performance.
As to the second part of this saying -- stripping the algorithms down brings its own pitfalls. Of course, not every strategy has to be sophisticated. For example, all volume weighted average price (VWAP) strategies from different brokers are highly similar: historic intraday volume distribution and price histories are combined to state and follow a simple collection of trading rules. But stating these rules too rigidly makes a strategy highly transparent, i.e. “readable”. Though most of the market participants take care to introduce some degree of randomisation and make the strategies “opaque” (or at least claim so), there is still plenty of evidence of naive algorithms at work.
Since many strategies formulate their targets in terms of 15 or 30 minute bins, some of them accelerate the pace of trading near the end of the bin, to avoid falling short of the target.
Figures 1 and 2 ( see download file) serve to demonstrate these artifacts. Figure 1 shows an average (normalised) number of trades for each 10-second interval for the 500 largest NYSE stocks. One can easily see the greatly increased trading activity every 30 minutes, somewhat smaller increases every 15 minutes, and yet smaller spikes every five minutes – the latter is less noticeable, but can be determined by the means of spectral analysis.
Figure 2 shows a similar picture for intraday price volatility, once again normalised and averaged for 500 NYSE stocks over the course of 2004. Similar results can be seen for Nasdaq securities.
These effects have become more pronounced in the last year or two with the wider spread of algorithmic trading. Other market artifacts, which have arisen more recently, can also be attributed to the automated strategies which are programmed to operate in a rigid way.
The details of a trading algorithm’s implementation are key factors to performance. A trading strategy that will be successful in the long run should be sufficiently robust against a variety of trading conditions and offer predictable price performance versus the benchmark.
Optimally, a strategy leverages the discipline of a mathematical model to assist in defining an objective, and in this sphere employs trading heuristics to achieve those goals. This is most likely where most firms seek to differentiate themselves from their competition, and unfortunately where the most “hype” appears in the documentation as well.
Trading heuristics as defined here may include the judicious use of statistical measures such as projected volatility or volume to help determine the utility of completing a trade at a specific price level to attain the objectives.
Methodology
A typical VWAP strategy derives its trading schedule from a historical volume distribution as a function of time and methodically executes the fraction of shares required to meet a target completion rate. The natural progression is from passive orders replaced with aggressive orders after some elapsed time, which may be predetermined or randomised.
Inevitably this causes the strategy to accrue shares at the end of each time bin that need more urgent order types to avoid rolling into the next bin. As shown in the graphs, VWAP engines that use time intervals between five and thirty minutes can easily be gamed by specialists and other market participants.
Some engines allow one to relax the amount by which an order may lag the distribution by setting bands around it, the argument being that more passive orders will have less impact. Eventually more aggressive orders will have to be applied to the accrued shares in order to maintain control – likely at adverse prices.
However, there is a limit to how much a VWAP trading strategy may adjust its schedule. Imagine a stock is trading at what seems to be the high of the day at 11:30am; if the strategy decides to accelerate selling at the current price levels, expecting the price to revert, it will be exposed to the risk that: (a) the price remains the same but trades on excessive volume at the close or (b) the price continues to rise and trades on normal or excessive volume at the close.
Alternatively, if at 11:30am the price seems depressed and the strategy decelerates trading on the expectation that the prices will increase, it will be exposed to the risk that the prices may continue to decline.
More importantly, there is an implicit compression of shares into the remainder of the day and therefore higher impact. The appearance of volatility spikes is likely an artifact of the way most time slicing engines are implemented.
For illiquid securities time slicing is out of the question – the significance of a volume curve is low, daily volumes are erratic, spreads are wide and trades infrequent. At least three order behaviours are responsible for achieving this: one to participate passively in the first tier of liquidity with sufficient randomisation, a second to capture unexpected price volatility, and a third to locate additional pockets of liquidity.
Here, as with implementation shortfall strategies, the ability to locate and identify liquidity critical to price improvement for a level of risk aversion is highly dependent on the quality of the algorithm employed.
When comparing providers of algorithmic trading it is important to look for high quality data sources, robustness of statistical techniques involved and a deep understanding of the marketplace and trading. One must also to look for a provider who is committed to algorithmic trading and able to devote resources to ensure their algorithms stay ahead of the curve as the markets change and as market microstructure evolves.
For more information please contact
The Miletus Trading Research Department
at research@miletustrading.com
