Systematic volume spikes and intraday liquidity patterns:

Fingerprints of HFT activity?

Discussion by Katya Malinova

Will J. Armstrong, Laura Cardella, and Nasim Sabah

Main Results

"Non-macro-news days" still have "teeth"

Volume "clusters" at discrete time nodes

Only partially explained by news anouncements

  • Conjecture:
    • "volume spikes" driven by HFT rebalancing activity
  • \(\Rightarrow\) Propose a new HFT proxy: 
    • use the "spikes" as a proxy for HFT liquidity provision

Main Results

Key Insight: Trades Cluster!

No such thing as "steady trade/order flow

\(\Rightarrow\) Caution is needed when interpreting traditional (esp. time-weighted) measures

  • Malinova and Park (2019 WP):
    • 67% of dollar volume in "within-10ms-clusters"
  • Admati&Pfleiderer (1988)
  • Easley, Lopez De Prado, O'Hara (RFS 2012):
    • "The most important aspect of high frequency modeling is that trades are not equally spaced in terms of time."
    • Mandelbrot and Taylor (1967): true even in 1960s

Key New Insight

Trades Cluster at Discrete Nodes

Volume is (predictably?) higher at 10:00, 10:30, 11:00, 11:30,...

  • Broussard and Nikiforov (2014; "Intraday periodicity in algorithmic trading"; Journal of International Financial Markets, Institutions and Money) document similar patterns
    • ​30-second buckets

Broussard and Nikiforov (2014): Volume

Broussard and Nikiforov: % Changes in Volume

Question: What causes the spikes?

The authors:

  • "almost surely machine based" \(\Rightarrow\) HFT
  • + improvements in liquidity
  • \(\Rightarrow\) Conjecture:
    • HFT aggressively rebalancing their inventories 
    • Coordinate on discrete points to "pool liquidity"

Alternate explanations?

Question: What causes the spikes?

Broussard and Nikiforov (2014):

  1. a result of optimizing behavior by algorithms that seek to concentrate their trading. 
  2. a side effect of humans’ (programmers) tendency to prefer round numbers. 
  • Third alternative = the combination of the two:
    • only some algorithms have rigid round time marks
    • influence the behavior of “smarter” algorithms that do not have such constraints.
  • They claim to support #2 but not #1.

Source: Doug Clark (ITG Canada); TSX Data on November 30 2011

Orders submission periodicity (sub-second)

  • Buy-side algos split orders across time
  • These algos may revisit the mkt every few ms/sec/min
  • \(\Rightarrow\) trading at predictable times and (?) predictable quantity

Likely "machine based" -- but need not be HFT

  • \(\Rightarrow\) may well be HFT trading during spikes
  • but not necessarily because of rebalancing inventory

Question: What causes the spikes?

Question: What causes the spikes?

Suppose buy-side algos trade at discrete nodes:

  • If aggressive \(\Rightarrow\) HFT may post aggressively in anticipation
  • Or: if buy-side uses aggressive limit orders, HFTs may trade against these passive orders

Can we identify/disentangle cause and consequence?

\(\Rightarrow\)  improvements in the spreads possible in either of these

Other features may also match the authors' findings, e.g. more prominent/easier to detect in small stocks

Measure of HFT Liquidity Provision?

What about the costs?

If HFTs respond to predictable buy-side trading:

  • can still be viewed as "providing liquidity"
  • Irrespective of whether HFT is passive or active
  • \(\Rightarrow\) Volume at the "nodes" may correlate with HFT liquidity provision
    • But: volume spikes may also be present with little HFT participation ...

Dig deeper into "spikes" & liquidity?

If HFTs "front-run" predictable, say, buy orders

  • cancel/take out the asks, "run the price up"
  • sell "high" and then buy as the price reverts to "normal"

Spreads can be higher but what about price changes? 

  • Do you observe price changes around the "nodes"?
  • Before or after the "nodes"?
  • If "front-running" \(\Rightarrow\) the price impact of trades at the nodes may be low or even negative -- yet a cost!
  • Caution when dealing with events:
    • Can event-related spikes driven by a single event (i.e. major unexpected news on one particular day)?

Other Comments

  • More explanations, precise regression equations, robustness? 
    • Timeframe? 1 second is not that short ....
    • Intraday Amihud: correlation with other intraday measures? 
    • Economic magnitude  (% spread \(\times 10^6\) in bps?)

Filtering the announcements:

  • Does Ravenpack do it or did you do this manually?
  • What is news is delayed?

@katyamalinova

malinovk@mcmaster.ca

slides.com/kmalinova

https://sites.google.com/site/katyamalinova/