Taming the Wild Algos

"And now," cried Max, "let the wild rumpus start!"

— Maurice Sendak: Where the Wild Things Are

It’s not just equities and futures markets where strange stuff happens! An “algo gone wild” was spotted in the oil market (it actually happened earlier this year) and intrepid Reuters journalists got to the bottom of it.

High frequency trading firm Infinium Capital Management is at the center of a six-month probe by CME Group (and reportedly the CFTC) into why its brand new trading program malfunctioned and racked up a million-dollar loss in about a second, just before markets closed on Feb. 3. The algorithm, which was brand new, went live 4 minutes before the end of trading. It fired in 2000-3000 orders per second before being shut off. The oil price surged $1 then slid $5 over the course of the next two days. Read about the full story here:

http://www.reuters.com/article/idUSTRE67O2QQ20100825

I know the CEO of Infinium Chuck Whitman from the CFTC technology advisory committee – he’s a good guy and very knowledgeable. I believe him when he says his wild algos had no malicious intent – the algos were just broken and shouldn’t have been put live.

With algorithms and HFT comes the possibility of mistakes. Many more firms outside of the equities world are embracing HFT and their inexperience can cause market disruptions such as the Feb 3rd CME issue. A flash crash in oil or other commodities - or even foreign exchange - is not to be scoffed at. In fact, many commodities markets are much less liquid and homogenous than equities, and can be even more vulnerable to mistakes or manipulation. In the case of Infinium, the algo caused a spike in trading volumes by nearly eight times in less than a minute. It was a classic case of the algo running wild until it faltered and 'choked'. This is not how HFT strategies are supposed to work.

There are a number of best practices that can be used to mitigate against algos going wild:

The first best practice is diligent backtesting – using historic data and realistic simulation to ensure many possible scenarios have been accounted for. What does the algo do in a bull market, a bear market, at the open, at the close, when unexpected spikes occur, during a flash crash, when non-farm payrolls or other economic news is released etc. etc.? Of course there’s always the possibility of a “black swan” scenario – but then there’s always the possibility of an earthquake in London – but I bet the buildings aren’t built to withstand one – it’s a matter of covering likely possibilities as best you can. A backtesting process needs to be streamlined of course – as short time to market of new algos is key.

A second best practice is building a real-time risk firewall into your algo environment. Just like a network firewall stops anomalous network packets reaching your computer, so the risk firewall should stop anomalous trades getting to trading venues. These anomalous trades might be human or computer generated – such as “fat finger” errors, risk exposures (for a trader, a desk or an institution) being breached, or even algos gone wild (e.g. entering spurious loops and spitting out anomalous orders). Real-time risk monitoring is a second level protection for those problems you don’t catch in backtesting.

A third best practice is to use real-time market surveillance in your algo environment. Even if trades do not breach risk parameters, they may breach compliance rules, regulations or may be perceived by a regulator as market manipulation (by accident if not design). Detecting these patterns as they happen enables good internal policing by trading firms, rather than investigation or prosecution by regulators.

An algorithm is a tool in a trader's toolkit, and it needs to be taken care of as such. If it is well-oiled and the trader or quant or risk manager monitors its progress then the algo will do its job quickly and nicely. If the trader/quant/risk manager doesn’t properly prepare the algo or ignores the algo and lets it get rusty, so to speak, it could lose its edge and run amok. Algorithms must be monitored constantly for performance and for errors, and sometimes tweaked on-the-fly to ensure best results. A good algorithmic trading platform will enable trading firms to do just that.

Trading firms are not the only ones who need to be on guard for possible algos gone wild. In the case of Infinium, the regulators and the exchange were also slow on the uptake. This shows that everyone needs to be proactive in using the correct tools to monitor algorithmic trading. Sensing and responding to market patterns before the aberrations or errors have a chance to move prices is the right thing to do - in all asset classes. Be like Max and tame the wild things!