The Gateway to Algorithmic and Automated Trading

Fit to Fade? News Driven Algorithmic Trading Strategies

With potentially terabytes of conflicting data arriving on multiple ultra low latency data feeds, a myriad of ways to slice and dice the flow of information, and the not inconsiderable challenge of accurately factoring in sentiment and expectation to challenge any news algo, automating trading from news and getting it right is possibly one of the greatest challenges for a high frequency trading firm. Bob Giffords analyses the extent and sophistication of some of the news driven trading strategies currently in operation.

The survey carried out last year by Automated Trader found that overall, news driven trading signals were used by around a fifth of the high frequency traders responding. Not surprisingly, a relatively higher number reported using event or machine-readable newsfeed data (15.2%) compared to those specifically using sentiment metrics or other news analytics (10.3%). Event data scored particularly highly in Asia, whereas no one in Asia was using sentiment type analytics, which were found mostly in North America or, to a lesser extent, Europe.

Automated news has already had a tangible impact on the markets. Macro data releases are now incorporated into FX and other prices in a couple of seconds [See "e-FX: Speed or Bleed"]. From the lock up rooms in global capitals, where announcements are made to waiting trading engines in London, New York, Frankfurt or Singapore takes mere milliseconds as high-speed data networks and electronic workflows spring into action. The old volatility and widening spreads have largely disappeared as markets can now absorb news and reset in much cleaner step functions unless of course there are significant surprises or uncertainties.

We normally think of machine-readable news as text processing, but of course news can be published in many formats. Increasingly scheduled economic news is released as tagged data and regulatory initiatives in the US and Europe will require more data releases and corporate actions of listed companies to be filed electronically, using perhaps XBRL or other formats and protocols. Those who wait for the text will then arrive late to the party unless they can capture a note of uncertainty or context that might change the view based on the raw numbers. The weird and wonderful world of robo-newsreaders has arrived.

However, the links between high frequency trading and newsfeeds do not stop there. Simply by counting stories and classifying them, risk advisors are now developing a range of news analytics to provide very short-term, even intraday, indicators of volatility and volume spikes. Directional signals are proving more elusive however. Nevertheless, recent research suggests that the well-known effect of volatility clustering may be related to the serial correlation of news releases. That would mean that news metrics could well add additional insight to GARCH models and implied volatility metrics based on options prices for example. Risk measures of course use different forecasting horizons and what applies for the next month cannot necessarily be used to drive decisions over seconds. Risk horizons and alpha horizons need to be aligned. Given the shorter alpha horizons of much trading, traders need to develop suitably adapted risk metrics, and news appears to offer this [See "Using News as a State Variable in Assessment of Financial Market Risk", Dan diBartolomeo, Northfield News, March 2010].

At the end of July the Financial Times noted that in English language media over the previous six months mentions of V-shaped recovery remained fairly constant, but talk of a double dip had soared, with pessimistic items rising four-fold since May. Consequently they argued that point risk estimates are of little help because of the wide range of potential scenarios and so the article called for more scenario-based forecasting [See "World Economy: Vulnerable to Verigo", Chris Giles, Financial Times, 27 July 2010]. Only by scanning the news can algorithms truly 'see' such scenarios unfold. The rest of us will only 'feel' the developments through our visceral appreciation of what we read, which is much less scientific and more instinctual.

Moreover, this example of growing macro-economic pessimism illustrates the layered semantics of most news. Quantitative news models definitely will need to take a layered approach focused more on context and scenarios than perhaps traditional technical pricing models have done. Institutional traders need to see any macro data release within several different contexts:

  • Analyst forecasts for the particular release, the historical accuracy of earlier forecasts and an estimate of how much of the consensus forecast had already been baked into the price.
  • The revision history of earlier releases.
  • The theoretical pricing impact of the data on its own.
  • The market context in which the announcement was made: does this announcement reinforce the trend or contradict it? The market appears to take more notice of contradictory messages. Are there other events that might override the normally expected market response? And finally....
  • The evolving market response itself and the weight of money behind it.

Deriving trading signals from news is never straightforward. Clearly any resulting model is likely to be highly complex and require an enormous amount of reference data, given the large number of macro data releases now scheduled. Automated methods become obligatory just to handle the volume of information and the sheer permutations and combinations of emerging results.

Similarly, the May 6 2010 Flash Crash showed us the impact that sudden shifts in sentiment can have on the markets even intraday [See "What Just Happened?"], whilst the previous month's volcano in Iceland gives another example where an exotic physical event had a very tangible market impact because of the political decision to close European air space. Had the volcano fully erupted there could also have been an impact on global crop yields. So traders who are using news feeds confirm that they scan the news for a wide range of signature events with cascading models to assess their cumulative impact or alerts to traders to intervene manually.

For traders, who use textual news to gauge market sentiment, there are further opportunities. For example, some traders are comparing the raw data feed with the subsequent sentiment of the related news reporting. This may corroborate or contradict the trader's own interpretation. The metric is then the 'distance' between the two views rather than the market sentiment measured on its own.

Other traders are analysing how the macro economic events are playing out on the fortunes of individual companies or sectors via ETF prices. Reference data describing supply chain flows across industry sectors can also be used to pick up unexpected impacts according to one commentator. Naturally all of this algorithmic infrastructure will potentially result in higher correlations between markets [See "Perfect Storm"].

There appear to be many fruitful avenues for investigation. Some researchers are trying to understand the feed through mechanisms for news to retail or institutional investors. It appears that the retail investor may be more influenced by the raw news directly, while the institutional investor may respond more to analyst opinion, so the focus then shifts to how they are influenced.

Since high frequency trading is still a relatively small niche in the market, only a tiny fraction of traders overall is probably using automated newsfeeds at the moment in their trading models. Software vendors who offer adaptors to the commercial feeds do not report heavy take-up, but usage is also clearly growing.

With so much correlation in the markets, portfolio managers are inevitably looking for diversification strategies to avoid the dangers of quantcentration effects [See 'News Analytics in Finance', by Professor Gautam Mitra, Leela Mitra, Optirisk Systems, based on research undertaken in CARISMA at Brunel University, John Wiley, 2011], where common data and model sources lead to crowded trades. News in all its forms seems able to offer that diversity, as an additional component in multifactor models. In taking the plunge it is important to decide the type of news automation required, since each involves quite different tools and skills. There appear to be four quite different approaches:

  • Low latency macro or micro data releases.
  • Sector or individual company news sentiment..
  • Macro news analytics.
  • Regime change early warning systems.

Low latency data releases are the most mature and widely deployed of the news technologies. Over the last couple of years they have been expanding to cover more types of macro-economic data and more global regions, like for example the Deutsche Boerse's AlphaFlash service. The availability of microeconomic data will grow as more companies publish their results electronically. The key issue here is time to market with the right information. However, since all market behaviours are linked ultimately to macro trends, traders may consider this an appropriate place to start in order to establish a reliable framework for assessing other text-based news. We cannot believe everything we read in the papers.

Sector, company and product sentiment is another potential candidate for serious investment, but involves very different tools and research databases. Dow Jones and Thomson Reuters have both invested heavily in these types of services and now have quite comprehensive second generation offerings with extensive back testing capabilities. A few brave hedge funds are trying to do their own textual analysis using their own reference data, language dictionaries and grammars. This allows them perhaps to apply different natural language processing or artificial intelligence techniques compared with the mainline commercial offerings. However, firms should not underestimate the challenge and cost of going it alone.

The third type of strategy might be described as macro news analytics, which are really metadata about the newsflow itself, rather than information about the news content. News analytics will measure, for example, the frequency of different types of articles and compare them to some normalized rate in order to highlight extreme events and predict mean reversion. These metrics typically will not address individual companies or commodities, but instead will track macro-events like natural disasters, country political instability or sector developments like energy, agriculture or mining, etc.

The last strategy is focused more specifically on identifying anomalies that may signal regime change in market conditions. It may use all of the news types described above, but the objective is to provide early warning of those changes. Market makers may then respond by narrowing or widening spreads, execution algos may choose to become more or less aggressive, and systematic traders may similarly adjust their risk preferences and possibly rein in some exposures.

Over the past few years the volume of global news has exploded, and it tends to spike further when bad news predominates. As correlations across instruments and asset classes have increased along with volatility, it is becoming increasingly important to be aware of how news flow interacts with the markets. Newsflow technologies have similarly matured resulting in faster interactions between news and markets across regions and asset classes. If trading algorithms only look at the markets, they are essentially flying blind as storm clouds bear down upon them. Like radar, the robo-newsreader can help to navigate the storm.