Volume is no longer a measure of stock market liquidity

Reading the CFTC-SEC preliminary report on the crash of 2:45, I think we can draw at least one very clear conclusion:  In a world with high frequency trading, volume is not a measure of stock market liquidity.

Why?  Because the report states very clearly that volume was at its highest when prices were at their lows.  (See figures 13, 29, 30.)  In other words, the data doesn’t support the view that the high-frequency traders left the market en masse, only that they dropped their bids by more than 5% in a matter of minutes.

i.  What happened in the e-mini market

In particular on the e-mini market, a 2.5% decline from 1097 to 1069 took place over about one minute in an orderly market with unusually high volumes of trade and at normal bid-offer spreads.  Since the best spread throughout this process was less than 0.025% of the price, the speed of the price adjustments indicates that algorithms were very active in the market over this one-minute time period.

At 2:45:27 the price of the contract fell a full percent in half a second to 1056 and the bid offer spread gapped to more than half a percent of the value of the contract.  These events triggered a five second pause in the market — and the price recovered from there.

This series of trades successfully removed an existing (and persistent) imbalance in the limit order book that had favored sellers.  Thus, the algorithms were merrily trading away in this environment, until at 2:45:27 something — perhaps the accelerating drop in the market, perhaps the reversal of the limit order imbalance, or perhaps human intervention — led to a gap in the bid-offer spread.

It is interesting that this gap in the bid-offer spread — which would appear to be a sign of a disorderly market — was the harbinger of the recovery of the market.  It really only took one minute of  high volume trade at extreme bid-offer spreads, before the e-mini market returned to a state of perfectly normal liquidity — with trading volumes declining to normal as prices rose and there was a return to balance of the limit order book.

In short the CFTC data makes it appear that the algorithms trading on the e-mini market worked very efficiently to eliminate a persistent limit order book imbalance by driving the price of the contract down until the book was balanced again.

It is interesting that the CME’s analysis of the crash concludes:

However, there is no visible support of the notion that algorithmic trading models deployed in the context of stock index futures traded on CME Group exchanges caused the market fluctuations in question.

Rather, we believe that automated trading contributes to market efficiencies, generally bolsters liquidity and thereby contributes to the price discovery function served by futures markets. This view is supported in the academic literature where one study found that “the move to screen trading strengthens the simultaneity of price discovery in the cash and futures markets and lessens the existence of a lead-lag relationship.”4 Another study concluded that their “results are consistent with the hypothesis that screen trading accelerates the price discovery process.”5

Further, we find no evidence in CME stock index futures of any undue concentration of activity amongst algorithmic or any other types of traders. In fact, activity levels amongst various CME constituencies on May 6th were quite consistent with normally observed patterns.

Trading by the most active of these traders was generally balanced between buy and sell orders during the period from 13:30 to 14:00 (CT). It is difficult to attribute the declining market action to any concentration of high frequency traders. Rather, we suggest that HFTs may have had the effect of providing a buoyant function in the market.

I would be interested in seeing the CME’s second by second analysis of the trades that took place from 2:44 to 2:46.  It is simply hard to believe that humans could have managed a 2.5% one minute drop in an orderly market.  It also seems to me that market participants should do their best to leave their “beliefs” out of their analysis and avoid citing academic studies that are inherently ambiguous in their conclusions.

The last paragraph addresses concerns that are only peripheral to the issue of algorithmic involvement in the crash.  Since high frequency algorithms are designed to buy and sell with extraordinary speed, we would not expect the crash to result in an imbalance between the algorithms’ buy and sell orders — presumably they were buying and selling all the way down and all the way up.  The only point at which one would expect a small concentration of algorithmic activities is in the minute where the market fell from 1097 to 1069 — it seems likely that many human traders would have paused to evaluate the market activity during this period and thus that a disproportionate share of the non-algorithmic trades would be outstanding limit or stop loss orders.  Finally, I don’t think anyone disputes the view that algorithms were healthy participants in the recovery of the market from its low and in this sense buoyed the market.

In short, despite the CME’s argument that there is no support for the view that algorithms caused the fluctuation, the orderly speed of the drop in the market belies this claim.  Until the CME releases a second by second analysis of trades involved in the descent, it seems to me that we must lean towards the common-sense view that only computers could manage such an orderly collapse.  It seems that the CFTC and SEC agree with me.

ii.  Why volume is irrelevant

Between 2:45:19 and 2:45:29 17000 emini contracts traded at an average price of 1056 and in a 10 second interval approaching 2:49 about 13000 contracts traded at a price of about 1085.  Both these volume figures are well over 10x typical trading volume.  That is prices collapsed and recovered in a period of about four minutes on record trading volume.

The flash crash is a clear indicator that trade now takes place over such small increments of time that measures of volume have no meaning.  By dividing time into infinitesimally small increments, modern markets have created an environment where it is extremely difficult for long-term investor-buyers to trade simultaneously with investor-sellers.  Instead at the particular microsecond in which the investor’s order hits the market the only other traders available with which one’s order can be matched are high frequency traders.  Any imbalances between buyers and sellers that last for a second or more will be resolved by changes in prices.

What we observe in the e-mini market on May 6 was an imbalance that lasted for several minutes — and was very efficiently resolved by the algorithms trading on the market.  This is presumably just a magnified version of the micro-sized process that goes on all the time in our algorithm-driven markets.

The way to think about the relationship between volume and trading intervals seems to me to be the following:  every time the trading interval over which prices are set shrinks by an order of magnitude, volume must increase by an order of magnitude in order to maintain the same effective level of liquidity as before.  Since the trading interval has been shrinking much faster than volume has been growing, we end up with markets that are very thin, despite unprecedented trading volume — and on these “thin” markets a whole days worth of trading volatility can take place in a matter of minutes.

I suspect that unless regulators impose a limit on technology’s ability to fragment markets over time, “flash crashes” will become a periodic regularity in our markets.  While finding ways to slow down trade in individual stocks is useful, if it is in fact the case that it was algorithmic trading that triggered the market dislocations of May 6, then the solution to the problem will lie in the regulation of algorithmic trading, not the underlying stocks.