Francis Breedon, Louisa Chen, Angelo Ranaldo and Nicholas Vause
Most academic studies find that algorithmic trading improves the quality of financial markets in normal times by boosting market liquidity (so larger trades can be executed more quickly at lower cost) and enhancing price efficiency (so market prices better reflect all value-relevant information). But what about in times of market stress? In a recent paper looking at the removal of the Swiss franc cap, we find that algorithmic trading provided less liquidity than usual, at worse prices, and that its contribution to efficient pricing dropped to near zero. Market quality benefits from a diversity of participants pursuing different trading strategies, but it seems this was undermined in this episode by commonalities in the way algorithms responded.
Our analysis is based on 100-millisecond-apart spot-market quotes and transactions in several currency-pairs on the EBS Market platform, which were kindly provided to the Bank by NEX Markets. These data identify whether quotes were supplied and transactions executed by manual interfaces (MIs) or algorithmic interfaces (AIs). The latter are further divided into AIs operated by banks and those run by the professional trading community (PTC), which includes hedge funds and proprietary trading firms. Other than the type of interface used, the data contain no information about the characteristics or identities of platform participants. The data we received cover a period of three weeks centred on 15 January 2015. At 9:30 GMT on this date, the Swiss National Bank (SNB) unexpectedly abolished its policy of capping the value of the Swiss franc (CHF) against the euro (EUR), prompting it to appreciate by over 30% within minutes of the announcement. A related post examines the reaction in derivative markets to the same event.
For our analysis of market liquidity, we study the volume of trades for which each type of interface demanded immediacy (‘consumed liquidity’) by submitting an immediate-or-cancel (IOC) order and supplied immediacy (‘provided liquidity’) by providing a good-till-cancelled (GTC) order that was matched with an another order. In particular, we investigate the extent to which each type of interface was a net consumer or provider of liquidity, respectively meaning they demanded immediacy in a larger or smaller volume of trades than in which they provided it. We also examine effective spreads for each type of interface. These are differences within very short time windows between the prices at which currency-pairs were bought and sold.
Focussing on the EUR/CHF market, we find that AIs operated by PTC firms were net consumers of liquidity on the day of the SNB announcement, while MIs were net providers of it (Chart 1). The latter remained the case even after deducting from MI volumes estimates we made of SNB trading activity. In addition, effective spreads between trades in which AIs did still provide liquidity widened by more on the day of the announcement, especially for PTC firms, than for trades in which MIs provided liquidity (Chart 2). Statistical tests, reported in our working paper, show these patterns were significantly different to those of the preceding and succeeding periods. The less than complete reversal of effective spreads towards pre-event levels in the post-event period probably reflects the increased volatility of EUR/CHF following the SNB policy change.
Chart 1: Cumulative net liquidity provision
Chart 2: Median effective spreads
For our analysis of price efficiency, we adapt a commonly-used methodology to estimate changes in the prices of currency-pairs that reflect new value-relevant information, decomposing these into contributions from different sources. In particular, we estimate contributions from public information, which is embedded in past price movements, and contributions from information that is private to MI, bank AI and PTC AI traders, which is embedded in their order flows. We also examine the frequency and size of arbitrage opportunities within trios of currency-pairs and investigate whether each type of interface traded to profit from these opportunities when they were present. For instance, there would be an arbitrage opportunity if it were possible to simultaneously convert euros to Swiss francs, Swiss francs to US dollars (USD) and US dollars back to euros and, in doing so, make a profit.
Again focussing on results for the EUR/CHF market, we find the contribution of PTC AI order flows to changes in the efficient price declined to close to zero on the day of the SNB announcement from almost one-third previously (Chart 3). Although the share of total order volume from PTC AIs dropped sharply in the several minutes following the announcement, it subsequently recovered to pre-announcements levels, so this result cannot simply be explained by low trading. Separate but related analysis in our working paper shows that, at the same time, bank AIs contributed materially to additional price movements not consistent with new information. As they often did this while buying an appreciating Swiss franc, we speculate that it may have reflected automatic hedging of option positions. Arbitrage opportunities within the EUR/CHF-USD/CHF-EUR/USD triangle also became 100 times more frequent and 20 times larger on average (Chart 4). However, trading activity did not intensify for any trader type when these opportunities appeared, whereas it did previously, most strongly for PTC AIs.
Chart 3: Contributions to efficient price variance
Chart 4: Frequency of arbitrage opportunities
For non-CHF currency pairs, we find no significant changes on the day of the SNB announcement in the net supply or cost of liquidity for any type of trading interface. Some small arbitrage opportunities did appear, but these were traded against and had disappeared by the next day. Overall, we find little evidence that the reaction of algorithmic trading in the directly affected market spread to other markets on this occasion.
It is difficult to draw general conclusions from one case study. However, our results suggest that market quality benefits from having a diversity of market participants, who pursue different trading strategies not only in normal times but also in times of stress. Hence, if the share of algorithmic trading (which was 60% of the EUR/CHF market in our sample period) continues to rise at the expense of manual activity, future financial market shocks could bring even greater declines in market quality. One possibility that could offset this is that further development of algorithms, including as they experience more stress events, leads to their behaviour in such events becoming less generic.
Relatedly, the Financial Conduct Agency has recently reported on its supervisory focus around the development of trading algorithms, highlighting examples of good and poor practice, and the Prudential Regulatory Authority is consulting on a draft supervisory statement that would enforce further requirements in this area. The latter would include identifying senior managers responsible for pre-deployment testing of the behaviour of algorithms, including in stress scenarios, as they are created and developed.
Francis Breedon is at the University of London, Louisa Chen is at the University of Sussex, Angelo Ranaldo is from the University of St. Gallen and Nicholas Vause work in the Bank’s Capital Markets Division.
If you want to get in touch, please email us at firstname.lastname@example.org or leave a comment below.
Comments will only appear once approved by a moderator, and are only published where a full name is supplied.Bank Underground is a blog for Bank of England staff to share views that challenge – or support – prevailing policy orthodoxies. The views expressed here are those of the authors, and are not necessarily those of the Bank of England, or its policy committees.