MOO/MOC orders market impact

sharpe_machine · Post by **sharpe_machine** » Sat Sep 15, 2018 12:55 am

Imagine that I am going to trade intraday strategy, which enters stock XYZ at the open and exits at the close (by placing MOO and MOC orders).

1. Which percentage of average daily volume do you think it is possible to trade in such a setting?

2. Any suggestions on how to measure market impact in this case?

edit1: it is assumed that the universe is SP500.

anonq · Post by **anonq** » Sun Sep 16, 2018 11:38 pm

Almgren-Chriss market impact is close enough to model market impact on the close. I have access to imbalance data from nasdaq/nyse and have created my own model in order to be more precise which gives very similar results to the ITG paper. At times an exceptionally large percentage of adv is traded at the close but that's not really a relevant question. The max percent of adv one should trade should the result of an optimization where you maximize your expected return so in the simplest case maximize ( expected return - trade size * tcost - trade size * market impact) where market impact would be a function of your trade size among other things.

https://www.itg.com/assets/ITG-Trading-Around-The-Close.pdf

ronin · Post by **ronin** » Mon Sep 17, 2018 1:54 pm

> Which percentage of average daily volume do you think it is possible to trade in such a setting?

If you strictly trade only the opening and closing auctions, you don't really care about the adv of continuous trading.

And it's not really about percentage of adv anyway, it is about percentage of realised volume. The impact is convex, so you do care about the difference.

What else can anybody tell you? Everybody agrees that impact is linear for small sizes, and increasingly convex as your size goes up. Some people use geometric instantaneous impact, some people use exponential instantaneous impact. Not that you will see much practical difference.

The answer to how much you can trade depends on the returns and the volatility of your strategy. If you are only returning a couple of bp, you will run serious into problems before 1% participation.

goldorak · Post by **goldorak** » Mon Sep 17, 2018 4:21 pm

And just to make sure, you are of course not using "O" and "C" from OHLC data as open and close auction prices?

sharpe_machine · Post by **sharpe_machine** » Mon Sep 17, 2018 4:39 pm

@goldorak, No, I use auction prices and NYSE/NASDAQ imbalances.

sharpe_machine · Post by **sharpe_machine** » Tue Sep 18, 2018 5:14 pm

@anonq, @ronin Actually, I followed the following path.

I used squared-root impact for my backtesting. Approximately it is

IMPACT = const1 * spread + const2 * morning/evening volatility * np.sqrt(my_vol / adv).

And my backtests are pretty good with my_vol <= 0.001 * adv (given that I trade each symbol only once per day).

Currently, I trade at much fewer portions of that (approx. 0.0005 * adv), but it looks like that my alpha decays. Therefore, my impact model is not good.

So the question was how to design an experiment to collect some data from which I can infer better impact model and not to risk a lot of money?

edit1: it is pretty clear that we can do some A/B testing with random splits of the portfolio (or even days), but it may cost some not-close-to-zero money to run.