Algorithms doing their thing.

Photographer: Martin Leissl

Algorithms Had Themselves a Treasury Flash Crash

Matt Levine is a Bloomberg View columnist. He was an editor of Dealbreaker, an investment banker at Goldman Sachs, a mergers and acquisitions lawyer at Wachtell, Lipton, Rosen & Katz and a clerk for the U.S. Court of Appeals for the Third Circuit.
Read More.
a | A

On Monday, U.S. regulators released a long shrug of a report on the Great Treasury Flash Crash of Oct. 15, 2014, when the prices of Treasury bonds shot up between 9:33 and 9:39 a.m. and then shot right back down between 9:39 and 9:45.  The regulators don't know what caused the flash crash. But they do know what stopped it. This is maybe the most emphatic causal claim in the report

Around 9:39 ET, the sudden visibility of certain sell limit orders in the futures market seemed to have coincided with the reversal in prices. Recall that only 10 levels of order prices above and below the best bid and ask price are visible to futures market participants. Around 9:39 ET, with prices still moving higher, a number of previously posted large sell orders suddenly became visible in the order book above the current 30-year futures price (as well as in smaller size in 10-year futures). The sudden visibility of these sell orders significantly shifted the visible order imbalance in that contract, and it coincided with the beginning of the reversal of its price (the top of the price spike). Most of these limit orders were not executed, as the price did not rise to their levels.

People (algorithms) who trade Treasury futures can see all the bids and offers to buy and sell Treasury futures at prices close to the current price. During the flash crash, Treasury prices were going up because lots of people were trying to buy Treasuries and not so many people were trying to sell. But there were people (algorithms) who had placed orders to sell Treasury futures at pretty high prices, and as the price kept going up, everyone else discovered those orders. So the people (algorithms) who were frantically buying Treasury futures realized that there were other people who wanted to sell Treasury futures. Those other people never actually sold those futures -- their mere existence, glowering in the order book 10 levels away, was enough to scare off the buyers and turn them into sellers, reversing the flash crash.  

There is an obvious dumbness to this: The algorithms stopped their orgy of buying not because they got some new economic data, or because a new buyer spotted value and jumped into the market, but just because they saw their shadow and got spooked. (Spooked, not spoofed, though of course this reliance on order-book data is what makes spoofing work, and spoofing and flash crashes have some history together.) But that dumbness was a good thing. It made the crash self-correcting. We -- and the regulators -- don't know what set the algorithms off on their buying spree, but a reasonable guess would be that, whatever it was, it was dumb. There was no big news.  There was no big fundamental buyer.  Algorithms did their dumb algorithm thing, interacted with each other in ways that baffled and alarmed their human creators, and then worked it all out on their own.

So ... how big a deal was this? I mean, the move was was famously a seven or eight standard deviation event, one "that is supposed to happen only once in every 3 billion years or so," though to be fair one that's actually happened four times in the last seven years.  And it clearly had some psychological effect; consider the approximately 3 billion articles about Treasury market liquidity that have been published since the crash. (Or consider today's joint staff report, which doesn't quite explain the flash crash but dutifully addresses Treasury market liquidity worries.)

On the other hand, it's just a flash crash. I once said:

People hate flash crashes, and obviously they cause some people to lose money, but they have always struck me as sort of non-systemic, a technical glitch rather than a major fear. A sharp permanent drop in asset prices is scary. A sharp temporary drop in asset prices is kind of funny, honestly. 

Same for sharp temporary rises in asset prices, though you'd be hard pressed to find any laugh lines in this report. But the report does shed some light on who lost money in the Treasury flash crash -- and on who didn't. The big worry with flash crashes is that people will unknowingly buy at the "wrong" price: Unsophisticated investors will put in market buy orders and find that they bought a $35 stock for $99,999, or put in stop-loss orders and find that they sold a $35 stock for a penny. But in the Treasury market you don't get a ton of unsophisticated investors, and market and stop-loss orders don't seem to have had much to do with the flash crash: The regulators found "only a limited number of market orders" on Oct. 15, none of which "appear to have resulted in sizeable changes in prices," and "very limited" volume from stop-loss orders, which also "did not prompt sizeable changes in prices."

Instead, the people who were dumbly buying on the way up were more or less the people you'd expect: algorithms. Here's who was actively buying Treasuries:

Source: Joint Staff Report

The red bars are for "principal trading firms," that is, algorithmic high-frequency traders. As the price went up, they kept bidding up the price. As the price went down, they kept selling. They chased momentum, buying on the way up and selling on the way down. But don't feel too bad for them. Here's who they were buying from:

Source: Joint Staff Report

The red bars are the HFTs again. High-frequency traders were doing most of the buying on the way up, and selling on the way down, as active momentum-chasers. They were also doing most of the selling on the way up, and buying on the way down, as passive market-makers.  The flash crash, in the cash Treasuries market, was mostly a story of high-frequency traders making money off of other high-frequency traders.

Actually it's even sillier than that. A chunk of it was high-frequency traders making money off themselves:

A second notable aspect of trading on October 15 was the heightened level of self-trading during portions of the event window. Self-trading, for the purpose of this report, is defined as a transaction in which the same entity takes both sides of the trade so that no change in beneficial ownership results.

At the peak of the flash crash, self-trading in the 10-year "reached 14.9 percent and 11.5 percent for cash and futures" volume, respectively, and was even higher in the 5-year. "The bulk of self-trading in cash and futures markets was observed among PTFs, perhaps due to the fact that such firms can run multiple separate trading algorithms simultaneously." The aggressive momentum-chasing algorithms were buying from the passive market-making algorithms on the way up, and selling to the passive algorithms on the way down, and 15 percent of the time the same firm was running both algorithms. No humans were made better or worse off by those trades, though I suppose that the algorithms on the wrong side might feel a hot flush of algorithmic shame for losing out to their algorithmic co-workers. 

The report has some other stuff, about the relative liquidity provision from high-frequency traders and bank dealers, and about Treasury liquidity and market structure more generally, none of which is particularly shocking. The basic story seems to be what you'd expect: an initial stimulus, a pullback in liquidity provision, a spike in prices, a reversal when the market realized how far it was from fundamental value. We just don't yet know what the initial stimulus was.

But what I like about the Treasury flash crash is just how convincingly the algorithms mimicked human folly. For six minutes one morning in October, some computers built themselves a bubble. They bid up the prices of assets for no particular reason, just because all their algorithm friends were doing it too, and what algorithm would want to sell when everyone else was buying? And then they saw some big sell orders that spooked them and made them realize that they were at the top. So for the next six minutes they busily popped their bubble, selling down to more or less where they started. They did most of the work themselves: Algorithms bought from algorithms on the way up, and sold to algorithms on the way down. 

Should we fear our new electronic overlords for their control of the market? Should we pity them for their folly? I don't know. I will say, though, that a human could have missed the whole drama by going out for a well-timed cup of coffee. The algorithms eventually figured it out on their own. 

  1. So it's a weird sort of "crash," a term that usually connotes prices going down. But everyone calls it thatYields went down anyway. As the report puts it: 

    In the six minutes between 9:33 am ET and 9:39 am ET, the 10-year yield decreased 16 basis points. Between 9:39 am ET and 9:45 am ET, the 10-year yield then abruptly reversed course and nearly retraced the latter move, again with no apparent trigger. These sharp moves between 9:33 and 9:45 am ET represent the October 15 event window.

    Also, Treasury prices tend to go up on bad news, making the "crash" thing appropriate. Incidentally, I say "U.S. regulators" wrote the report because it's a joint staff report of the Treasury, the Federal Reserve Board, the New York Fed, the Securities and Exchange Commission and the Commodity Futures Trading Commission. 

  2. The report rejects a few possibilities, like a fat-finger or systems error (page 29), and mentions others in a dismissive way, like hedging of volatility bets:

    In particular, anecdotal commentary suggested that some dealers had absorbed a portion of the sizable “short volatility” position believed to have been previously maintained by large asset managers. As volatility spiked on October 15, those positions would have prompted some dealers to dynamically hedge this exposure, exacerbating the downward move in yields.

    That's on pages 18-19 and never mentioned again. The unwind of leveraged short rate positions also gets sort of an equivocal mention.

  3. Which is to say, it's not a causal claim at all: It just "seemed to have coincided with" the price drop. But you know what they mean.

  4. I assume, as the regulators seem to, that events in the cash market rapidly promulgated to the futures market and vice versa. The majority of activity in both markets, both on Oct. 15 and in general, is from what the regulators call "principal trading firms," what you'd probably call algorithmic high-frequency trading firms, and surely their algorithms let them arbitrage between cash and futures.

  5. I mean, a mildly disappointing retail-sales report came out an hour before the flash crash.

  6. I mean, "near the beginning of the event window, two buy market orders were executed in the 10-year futures market—one for 3,000 contracts at 9:33:45 and one for 2,100 contracts at 9:34:07—both of which coincided with reductions in market depth" (page 26). But $510 million worth of buying (each contract is $100,000 notional) isn't really huge considering that the 10-year future traded $779.5 billion dollars on Oct. 15 (table 3.2).

  7. See figure 2.5 of the report, which shows three other intraday moves that were bigger than the flash crash. Jamie Dimon

    Then on one day, October 15, 2014, Treasury securities moved 40 basis points, statistically 7 to 8 standard deviations – an unprecedented move – an event that is supposed to happen only once in every 3 billion years or so (the Treasury market has only been around for 200 years or so – of course, this should make you question statistics to begin with).

    Dimon, to be clear, wasn't actually saying, like, "Boy, we didn't expect this for another 3 billion years." Treasury rates are not normally distributed, etc.

  8. As the report says (page 23):

    The analysis suggests that PTFs and bank dealers were the main contributors to the pattern of net aggressive flows, consistent with their large share of overall trading volume, with PTFs accounting for much of the imbalance in aggressive flows during the event window across futures and cash markets. At the same time, there is strong evidence from a study of net passive flows to suggest that PTFs, as a group, also remained engaged as liquidity providers throughout the event window, thus pointing towards more than one type of PTF strategies at work.

    Footnote 19 explains the classification:

    Every trade by definition comprises an “aggressive” and a “passive” side with the “passive” defined to be the standing order to buy or sell an instrument in the order book, while the "aggressive" order is that which is executed when matched against a standing "passive" order.

    The figures in my text are from the cash market. In the futures market the story is more complicated (see figures 3.6 and 3.8 on pages 61-62 of the report), with HFTs and bank dealer desks actively buying (and hedge funds actively selling), while HFTs, dealers, hedge funds and others were passively selling (and asset managers were passively buying). 

  9. Page 28 is interesting:

    Although they significantly reduced their depth of orders, the data also show that PTFs as a group continued to provide the majority of order book depth and a tight spread between bid and ask prices throughout the day, even during the event window (Figures 3.23 and 3.24). In contrast, during the event window, the bank-dealers that remained present in the market significantly widened their bid-ask spreads such that they only provided limit orders at a substantial distance from the top of the book.

    In very broad terms, therefore, PTFs, as a group, reacted to the event of October 15 primarily by reducing limit order quantities, while the bank-dealers reacted by widening bid-ask spreads and, for brief periods of time, removing their offers to sell securities. 

    An analysis of the relative supply of liquidity in the bid and offer sides of the order books by participant type shows certain imbalances in the provision of liquidity during the event window (Figures 3.21 and 3.22). PTFs, as a group, contributed to the order book in a relatively balanced fashion throughout the window, often providing standing bids and offers of approximately similar sizes, though at low absolute levels. In contrast, the balance of bids and offers supplied by bank-dealers was considerably more variable during the event window. During the first part of the event window, as dealers intermittently removed their orders to sell securities in the cash market, the balance of their remaining orders became skewed toward purchase orders. Despite limited and at times imbalanced order book participation, bank-dealers continued to trade actively during the event window, and indeed the absolute volume of their trading increased. 

    (Emphasis added.) Also worth noting is figure 3.15 (page 63), which shows the collapse in orders from algorithmic traders at the time of the flash crash.

This column does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.

To contact the author on this story:
Matt Levine at mlevine51@bloomberg.net

To contact the editor on this story:
Zara Kessler at zkessler@bloomberg.net