Forums  > Trading  > Estimating price impact in the opening auction  
     
Page 1 of 1
Display using:  

bs2167


Total Posts: 19
Joined: Feb 2010
 
Posted: 2017-02-08 19:58
I should start by stating that I trade crap (messy, illiquid listed US equities). So, while my typical trade size is modest in dollar terms, it can often be enough to cause price movement. I'm trying to properly quantify this impact for my MOO orders but there are a few things about my approach that I find troubling. Here's what I'm currently thinking:


1) For opening auctions where I did not participate, measure the average price difference between the official opening price and the 9:28am (opening orders deadline) mid price. I expect this to be zero.

diff = ln(openPx / mid928)

avgDiff = [diff(1) + diff(2) + ... + diff(n)] / n


2) Repeat step 1 for auctions that I did participate in and calculate the adverse movement. For simplicity, I'll assume all buy orders here.

if BUY order:

myDiff = ln(openPx / mid928)

adverseMove = myDiff - avgDiff


3) Run a regression against my order size (as % of ADV) and the adverseMoves to get an idea of the sensitivity to position size



-The first issue I have with this is that the bid/ask spreads at 9:28 for this illiquid stuff are often nonsensical. Garbage in, garbage out.

-The second problem I can see is that there is some alpha here so I'd expect some adverse drift from 9:28 until the open regardless of my order being present or not

-Lastly, step 3 should probably be normalized for volatility. I could either include historical vol in the regression or perhaps only compare each stock against itself when calculating adverse moves. The latter would be somewhat of a pain to manage.


Anyone have a helpful suggestion or two? Or perhaps I should be taking a different approach entirely?

Appreciate any input.


punx120


Total Posts: 1
Joined: Feb 2017
 
Posted: 2017-02-08 20:24
you could also look at the reversion after you trade, so open to 9:35, 9:45, etc... You could also try to remove the market move - but may be tricky as you said it's illiquid stuff...

also how much of the auction do you participate? you could also send LOO to protected yourself

EspressoLover


Total Posts: 205
Joined: Jan 2015
 
Posted: 2017-02-08 20:40
Why not do A/B testing? Randomly pick some small percent of symbol-days where trading is shut off. Then compare the regressions between the on-group and off-group. That removes the problem of distinguishing between your alpha and your impact.

bs2167


Total Posts: 19
Joined: Feb 2010
 
Posted: 2017-02-08 22:03
Thanks for the responses punx and EL...

To answer the size question, my desired level of participation is dependent on the price sensitivity. Currently, I limit it to .5% of ADV but that's likely too high. I read an ITG paper that has the average US MOO at 1.4% of daily volume (http://www.itg.com/thinking-article/measuring-the-trading-activity-at-the-open-and-close-actions-around-the-globe/). My figures have it a bit higher for smaller issues but it's very hit or miss.

Anyhow, to both of your points on methodology:

The A/B testing should allow for a clean benchmark against a 9:35 mid. That gets me most of the way there (thanks again) - just need to work through normalizing for volatility. Essentially, I'd like to come out of the exercise with the ability estimate the % impact on the opening print for a given level of volume. I suppose rather than get cute, I could just use a lookup grid with ADV% and HV% quintile buckets (or similar).
Previous Thread :: Next Thread 
Page 1 of 1