Improving VWAP Performance
This morning’s note is intended to give a cursory refresher course on how VWAP works, talk about how typically our industry has gone about trying to improve VWAP performance, and finally to tell you how we think the industry should go about trying to improve VWAP performance.
Volume Weighted Average Price (VWAP) institutional trading strategies date back at least to our time at Instinet in the 1990s. The stock market was hot back then, and foreign fund managers were at least as eager to buy US equities as taxi drivers and barbers. Overseas clients going home for the day would give their US stock orders to their US brokers each morning, and get their fills back from them the following morning.
Comparing the fills to the intraday charts was a sometimes sad and sometimes humorous exercise. You see, some of those US brokers took their best execution duties less seriously than they should have. As a matter of fact, the sarcastic slang terms “Really?” and “Are you F$%$ing Kidding Me?” and “WTF?” were first coined by London traders first receiving back their US execution fills T+1. Contests were actually had nightly in London pubs over who got the day’s worst fills. These London traders dreamed of a day where they could hope for a marginally average, and not-horrifically-bad, trade execution – and the VWAP benchmark was born.
Buyside firms have generally embraced the VWAP concept through the years. In fact some firms based their traders’ compensation on their executions versus VWAP. Slicey Dicey VWAP Algorithms were developed, and have become a staple in nearly all bulge firm algorithmic suites. Although these algorithms have varied sophistication, their general premise is to provide an average, or mediocre, execution. Traders embrace them because, while they will never “hit a home run”, they generally don’t strike out either – at least in large cap stocks if their volume participation is kept in check. As more traders execute trades using VWAP concepts, the day’s volume curve has increasingly become more predictable, and a self-fulfilling prophecy.
How Industry Tries to Improve VWAP
Industry attempts at studying VWAP algorithmic performance have centered on reducing the VWAP tracking error by trying to best predict the day’s volume correctly and getting “the smile” shape correct as well. Different firms do this in different ways. Some firms divide up the trading day in 5 minute buckets – 90 in total, and use historical data (20 day, 30 day, 3 months) to estimate the volume that will trade in each bucket. Other methodologies involve shorter, or even longer, bucket times.
For example, Flextrade published a report this summer titled Predicting Trading Volume and Volume Percentages. In that report Flextrade advocates using 26 15-minute buckets, as they believe that volume in those longer interval buckets is more accurately predicted. Their paper is quite good, and worthy of your reading; please download the paper at the above hyperlink.
Flextrade makes the case that VWAP execution performance can be improved by using not just historical volume data, but by using predicted bucket raw volume:
When predicting volume, convention is to reference historical averages as the base case. Relative to that base case we show improvements in the prediction of raw volume of 29%. In the case of volume percentages, we improved over the base case by 7%. More importantly we show that using our predicted volume percentages improves the performance of the VWAP algorithm, relative to historical averages, by 9%.
So, Flextrade can perhaps help you miss the VWAP by 3/10ths of a penny instead of 4/tenths of a penny. Or something like that…
But Wait… Maybe We Need To Rethink How We Look At VWAP!
There are many well-meaning attempts to try to help you drop fewer crumbs of cake when you trade. However, we submit to you that we are looking at it all wrong.
Consider these observations:
– VWAP algos do slice the parent order into child orders divided into time buckets for execution.
– Within those buckets (15 minutes long in some cases), stocks can trade in a wide price range (say between $40.10 and $40.15.
– VWAP algos are spotted easily by short term traders using various means (looking at the uptick in inverted exchange and ADF prints is one way).
– Short term traders, when spotting a VWAP algo, like the idea (let’s use our $40.10- $40.15 bucket price range) of buying between price points $40.10 – $40.13, and selling at price points $40.14 – $40.15.
– Broker dealer dark pools and SORs want to minimize routing costs in a big way, and so love the idea of short term traders “adding liquidity” to their pools.
– Some broker dealers likely have their own short term traders active in their pools.
– Broker dealers create VWAP algos for you to use that are absolutely cost sensitive; they are tooled to get rebates, and are tooled to avoid routing to high cost places.
ITG’s Maureen O’Hara wrote a paper in April 2014 titled High Frequency Market Microstructure. There is one section of the paper that we think stands out. O’Hara did a study of VWAP trades that were executed with a standard VWAP algorithm from ITG in 2013. She found:
“The sample size is 243,772 parent orders. The algorithm executed 13,468,847 child trades, meaning that on average each parent order turned into 55.325 child executions. The data also show that the algorithm executes the vast majority of parent orders with passive executions. For the sample as a whole, 65.3% of trades were passive; 21.9% were midpoint trades; and 12.57 % were aggressive. Less than one in eight executed trades actually cross the spread.
Some of you may be thinking that the ITG VWAP algorithm must be good, as it is executing “passively”. We would like you to possibly think about it in an alternative way. Let’s go back to our example where a stock varies in price in a time bucket between $40.10 and $40.15. Are Ohara’s findings consistent with the following? :
– Algo bids $40.10 on EDGX (high rebate exchange) instead of crossing spread at $40.11.
– Short term traders take at $40.11. Then bid at $40.11 and even $40.12.
– Algo joins bid at $40.12 on EDGX, and again does not cross spread and take at $40.13.
– Short term traders take at $40.13, and bid at $40.13 and $40.14. They also offer on EDGA at $40.15.
– Algo finally decides to bid at $40.14 on EDGX and 40.145 on EDGX hidden, and even cross the spread and take EDGA at $40.15.
– Short term trader sells on EDGA at $40.15, hits Algo’s $40.145 bid on EDGX hidden, and even $40.14 bid on EDGX.
To recap… the short term trader bought at $40.11-$40.13, and sold to Algo at $40.14-$40.15. The Algo paid up a multiple of pennies! Now, isn’t saving these pennies a much better use of problem-solving time than trying to save 1/10th of a penny? Aren’t these pennies bigger crumbs?
Imagine the short term trader acting as outlined above in response to every single VWAP child order.
If you can, then you can see that VWAP Algos are an alpha feeder. Who do you think is selling to you all throughout the running time of your algo? The sellers are not a cross section of retail, other algo orders, other investor resting orders. The sellers are a subset of market participants that have inserted themselves between you and the other “naturals”. They have done so not only in an un-needed way (these are liquid enough large caps keep in mind), but they have made you pay up. And they were enabled and encouraged either because your child orders fed alpha (brokers or third party), or because your execution interests were subordinate to the broker’s desire to minimize costs and garnish rebates. I think our industry knows all of this privately very well, even if they won’t talk about it on market structure panels and market structure notes.
We have a great idea for an academic study. Perhaps even ITG’s Maureen O’Hara can conduct it, as she already has the control case documented – the 2013 ITG Algo performance. Maybe ITG can create a new VWAP-style algorithm that is designed to always cross the spread first in each time bucket, and perhaps to do so at a random point in that time interval. Maybe this algo can be called Clean Weighted Average Price (CWAP), and its performance and tracking error can be contrasted with the control VWAP case. We wonder how that would turn out.
Maybe such a CWAP algo would make a fiercely-smiling algo into a happy-smiling one.