The Gateway to Algorithmic and Automated Trading

Baptism by Fire

Published in Automated Trader Magazine Issue 07 October 2007

Extreme volatility in the equity markets made August a testing month for both users of execution algorithms and automated trading strategies. Leading players explain the lessons for the future to Chris Hall.

Baptism by FireWhen fears of a collapse in the US sub-prime mortgage market quickly led to widespread panic in the financial markets in August, the finger of suspicion was soon pointing at some pretty familiar bogeymen. Hedge fund managers - aided and abetted by heartless, unseeing computerised trading models - were wreaking havoc, spreading contagion and putting granny's life savings at risk. Within days of BNP Paribas announcing on August 9 the suspension of three hedge funds over their sub-prime exposures, the financial pages were unanimous. 'Hedge Fund Panic Was Behind Global Stock Markets Collapse' screamed the UK's Daily Telegraph, 'Hedge Funds Prepare For Mass Redemptions', gloated the Financial Times.

As firms sought to establish and minimise their exposure to credit derivatives, liquidity dried up in the money markets and confidence in equity market valuations wobbled. As figure 1 shows, short-term interest-rate volatility rose persistently from mid-August as the extent of institutions' credit exposures gradually unraveled, falling away only when the US Federal Open Market Committee lowered its target for the federal funds rate 50 basis points to 4-3/4 per cent on September 18.

In the eyes of some, however, the financial markets would have veered much closer to collapse without recourse to technology-based trading tools to take logical trading decisions amid a period of sustained volatility rarely seen over the past decade. Large investment banks are still counting their losses from their exposure to the US housing market (latest estimates put institutional losses at $60-70 million), but neither banks nor their buy-side clients have been toppled by their use of computerised trading techniques. While, admittedly, some quant-based hedge funds and prop shops might still be nursing their wounds from over-reliance on statistical arbitrage models, buy-side users of execution algorithms may actually look back on August 2007 as a rite of passage from which many emerged more confident and aggressive users of algorithmic trading tools.

Fear, greed and algorithms

It could, however, have been very different. Had there been a repeat of July 7, 2005, when algorithms were ordered to be switched off by the London Stock Exchange in the wake of the London terrorist bombings, the equity markets would have been plunged into chaos, according to Brian Schwieger, Head of EMEA Quantitative Execution Desk, Merrill Lynch. "Following the LSE's request on 7/7, we saw a significant increase in volatility and spreads as the market lurched down," says Schwieger. "When the algos were turned off, the traders panicked because they didn't know how they were going to cope with the volume of orders. What causes volatility is and always has been fear and greed. In contrast, algorithms can take a logical statistical approach to execution. August 2007 has provided a counter-argument to the perceived wisdom that algorithms add to volatility."

"August 2007 has provided a counter-argument to the perceived wisdom that algorithms add to volatility."
Brian Schwieger, Merrill Lynch.


The equity markets witnessed a widening of spreads in the region of 20 per cent in mid-August compared to the previous three months, according to Merrill Lynch; far from the conditions many algorithms were designed to operate in. But rather than switching off their algorithms, many users simply changed to more aggressive tools designed specifically to capture liquidity in volatile markets. John Edge, European head of Electronic Client Solutions, JP Morgan, says that a clear understanding of the differences between algorithmic trading tools is crucial to riding out difficult market conditions. "The weight of responsibility must still be on the human trader to take the decision on when, how and with what tool to trade. This means understanding the values of various products, knowing when to change an algorithm, when it's not performing, and then choosing the right one for the circumstances," says Edge. Over the summer months, JP Morgan experienced an increase in the percentage of flows directed through its Arid and Aqua strategies. "These algorithms utilise the philosophy of high frequency trading and the demands of an active trader," says Edge. "Arid and Aqua are adapted from algorithms already used in the fragmented US markets, and have the ability to extract more liquidity than is readily apparent."

Increased Volatility in Price of Short-term Interest Rate Futures – April-October 2007

Increased Volatility in Price of Short-term Interest Rate Futures - April-October 2007

Merrill Lynch's Schweiger noted a "measurable" shift by clients from VWAP to implementation shortfall algorithms. While VWAP algorithms execute orders over a defined time period based on historic volume curves, implementation shortfall algorithms are more able to react to changes in order book volume because they do not have a defined end time. Although, like VWAPs, they also use historic market data, implementation shortfall algorithms can be calibrated to provide a more or less aggressive balance between alpha capture and market impact. Counter-intuitively, perhaps, Merrill Lynch's VWAP algorithm registered an improved performance in August. "You would expect VWAP algorithms to struggle because they're trying to predict the volume distribution, which in such circumstances is very hard," says Schwieger. "But this was more than compensated for by the way the algorithm reacted to the volatility."

Richard Balarkas
Richard Balarkas, Credit Suisse

"Opportunistic algorithms have performed even better than usual because volatility creates more of the opportunities that these strategies are designed to exploit."

Having reviewed the performance of different strategies, Merrill Lynch found that overall algorithms had been able to identify market volatility in August and adjust how they traded on the order book accordingly, resulting in a significant uptake of passive execution. The bank's OPL-X strategy - an aggressive implementation shortfall strategy that probes for hidden liquidity and reacts to increased volatility - registered a 40 per cent improvement in benchmark performance. "Recognising the volatility, the algorithm was able to play it passive and achieve some excellent results," says Schwieger. "Rather than chasing the market up, it was able to 'sense' the liquidity to cross and that allowed it to execute passively."

Horses for courses

Credit Suisse also observed an increase in algorithmic trading flow, notably from the bank's more opportunistic 'stealth' strategies, Guerrilla and Sniper, which are designed to source both displayed and dark liquidity with minimal market impact. According to Richard Balarkas, Head of Advanced Execution Services Sales, Credit Suisse, these algorithms and their compound versions now account for around 40 per cent of algorithm usage by clients compared with less than three per cent at the start of 2006. Following volatile market conditions in February 2007, caused by a large and unexpected stock sell-off in China, Credit Suisse analysed the relative performance of different algorithmic trading strategies. The bank's research concluded that stealth strategies outperformed against a VWAP benchmark by almost 70 per cent in high volatility conditions compared with normal conditions, while time-based strategies underperformed by 44 per cent. Interestingly, the relative performance of stealth strategies was 148 per greater in high volatility conditions with a strong downward direction than in a flat market with normal volatility. "Opportunistic algorithms have performed even better than usual because volatility creates more of the opportunities that these strategies are designed to exploit," says Balarkas. "In fact, they perform at their best when there is increased volatility and strong price momentum."

The level of support required by buy-side algo traders during August's market turbulence depended on their level of experience in using algorithms. "Firms that had been using algorithms for several years were very comfortable, but for others it was something of an awakening to the fact that each strategy has its own place and time, as they switched from VWAP to implementation shortfall algorithms," says Schwieger. There was also a marked increase in use of percentage of volume algorithms, which are well suited to unexpected volume spikes because they participate with order book volume at a rate defined by the user and tend to trade frequently and in small order sizes. "Historically, we've tried to steer clients away from using our percentage of volume algorithm on a regular basis. But when the market's moving rapidly in one direction and you want to be participating tick by tick, that's where it can really add value," says Schwieger.

Figure 2: Increasing Volatility in Stock Index Futures Prices - April-September 2007

But has the switch on use of execution strategies to take advantage of volatile conditions prompted a more permanent change in the algorithmic trading landscape? "Firms that started using algos in the summer haven't shied away, they've increased their usage. The August experience was a baptism by fire, an opportunity to get up the learning curve a lot faster than under normal market conditions," says Schwieger. "Similarly, the higher usage level of implementation algorithms seen in August has continued ever since. Increased awareness has led to increased usage."

John Edge
John Edge, JP Morgan

"There was a significant and unpredictable breakdown in the associations between the factors used in quantitative trading models."

Credit Suisse's Balarkas is also confident that clients will continue to take the more aggressive approach witnessed in August, and sees it as part of an overall trend toward greater sophistication in use of algorithms, characterised by increasingly frequent demands for customisation. Although clients can't re-engineer Credit Suisse's algorithms directly, they can request the bank to make changes at short notice. "We recognised that customisation was going to become increasingly important and increasingly complex," says Balarkas, "So we re-engineered our advanced execution services so that the tactics could be easily customised, and in some cases, changes can be made instantly. If, for example, a client asks for a live algorithm to be adjusted so that it acts more aggressively or leaves the residual in the closing auction, then we can react very quickly."

Quants trapped in rush for exit

For funds and prop shops using quant-based automated trading models, August's volatility presented a different challenge. Some funds reported a 'fire sale' of equity assets by hedge funds scrambling to source cash liquidity in the days immediately before August 9, in the expectation that the credit derivatives market was about to grind to a standstill. Advanced warning or not, trading conditions got tougher throughout August as market participants looked for safe havens from credit exposures. Figure 2 demonstrates the steady rise in the price volatility of leading stock index futures, reflecting the market conditions that shook correlations apart and widened spreads to make any exit a painful one. Like certain execution algorithms, the downside risk of volatile conditions for statistical arbitrage models in particular lay in their reliance on historical relationships that were no longer a useful guide for trading decisions in the prevailing market conditions.

"There was a significant and unpredictable breakdown in the associations between the factors used in quantitative trading models," says JP Morgan's Edge. "These models cannot be changed easily overnight and this gets exponentially harder intraday." This breakdown was compounded by the sheer volume of trading activity driven by too-similar trading strategies; the rush for the exit as funds tried to unwind positions made already difficult trading conditions that bit harder. "The widening of bid-offer spreads resulted from a lot of investors using the same strategies at the same time, particularly the quant funds and the long-short funds whose models use a multi-factor approach," says Matthew Carr, Head of International Equity Sales, BNP Paribas. "When they try to unwind large positions in one clip, this creates very high bid-offer spreads because everyone's looking for liquidity at the same time."

Amid rumours of quant funds losing up to 30 per cent of their value in August, turning off automated models was considered the sensible option as once-solid correlations irretrievably broke down. But hedge fund index compilers suggested an average monthly decline for August of around 1.3-1.6 per cent in early September (albeit based on a wider-than-usual dispersal of performances). And with the expected flood of redemptions failing to materialise, the overall picture is considerably more mixed.

"The jury's still out, for me," says Tom Heffernan, Director, Global Marketing at Last Atlantis Capital Management, a US-based alternative asset manager. "Some quant strategies that had a horrific opening in August came back strong by month's end," he says. "Our long equity programme, for instance, had a rough patch in early August, but bounced back well in September." Firms looking for early warning signs of an increase in volatility that could threaten models' assumptions were given mixed messages by the VIX, the Chicago Board Options Exchange's benchmark for stock market volatility. "VIX had been trading below 20 for such an extended period that many models likely weren't prepared for the new volatility spikes in August and September," says Heffernen. "Of course, VIX had been trending higher almost from the beginning of the year so those that stayed on top of that may have had an inkling a storm was brewing."

A model response?

According to Heffernen, longer-term trend-following models were particularly exposed. "Many historical trends have become increasingly short, if not imperceptible, and thus more susceptible to volatility swings," he says. "With the VIX fluctuating 30-40 per cent in one month, that's a tough environment in which a longer-term trend programme may deliver any value." But Heffernen believes firms could have been quicker to understand how a credit squeeze would impact prices. "When banks tightened access to leverage, managers were forced to exit the market at inopportune times. The lack of liquidity is a ferocious thing and costly, too. The experience has made everyone more cognisant of their correlation and exposure to the periphery," he says.

Matthew Carr of BNP Paribas
Matthew Carr of BNP Paribas

"Firms are assuming a major shock every five years and in some cases every other year."

Quantitative and high-frequency funds are already adapting their models in line with a risk profile that assumes higher levels of volatility. "History tells us that the kind of events that have a big impact on all markets come around every ten years or so. But given the speed of execution and broad access to markets that technology now affords, these situations may occur on a more frequent basis. Firms are assuming a major shock every five years and in some cases every other year," says Carr of BNP Paribas. "What firms need to estimate on the volatility curve is the gap risk; right now, most of the existing models assume a normally distributed curve. These models are very efficient under normal conditions, but when the markets are highly volatile, we might expect more a fat-tailed distribution, reflecting extreme market behaviour. Models must be more able to adapt within the less-normally distributed curve as fat tails or other similar outcomes occur." Carr says the difficulties of building models for volatile markets, given that backtesting is typically done against data from flat markets, are already being addressed. "Brokers are already looking at their algos and rebacktesting them taking August into account, and the large asset managers and quant funds should be doing exactly the same," he says.

Heffernan agrees that firms should be able to fine tune their approach to accommodate a wider range of market behaviour patterns. "You have to continuously evaluate your historical timeframe and focus your models on a more representative dataset. Does the current market volatility have a historical correlation? Does it have deviation symptoms similar to a time in the past? If so, those variables need to be invoked to where your model dynamically recognises the new conditions," he says.