Forums  > Trading  > Weird Trading Signals  
Page 1 of 1
Display using:  


Total Posts: 24
Joined: May 2020
Posted: 2020-11-27 14:31
I was backtesting some signals and found that if i go long on friday close and exit on monday close on an asset it generates a good sharpe.

If i include this signal in my current model the sharpe and returns both improve by around 10% .

What would be an appropriate way to treat this anamoly.Should i ignore it completely?


Total Posts: 474
Joined: Jan 2015
Posted: 2020-11-27 15:53
I don't think it's that weird of a signal. The weekend effect has been debated in academic finance for decades. It's also nearly indisputable that equity returns are disproportionately concentrated in the overnight period. There are pretty good economic justifications for daily/weekly seasonality effects. Due to psychological and institutional reasons, participants tend to demand different risk premias at different periods. A lot of people just plain prefer to de-risk to enjoy a less stressful weekend.

I think these types of signals tend to be underexploited. They're kind of an awkward fit for most of the major desks and strategy categories. A while back, I used to trade around the VIX term structure, and there were a lot of easily predictable daily seasonalities that persisted year after year. And that's one of the most studied/traded assets in the world.

All that being said, I'd still bet money that whatever you found is spurious correlation. Not because the hypothesis is infeasible. But just because almost any time a researcher tends to stumble upon an accidental hypothesis it ends up being spurious. IMO there's a huge latent state space of these potential "accidental hypotheses" that your mind is unconsciously and continuously scanning for when doing exploratory data analysis. You're not really even aware of the process, so even strong intuitive statisticians, tend to not correctly down-weight the multiple hypotheses effect.

I'd tend to be skeptical unless the t-stat is very large, or it's validated on totally out-sample data. Barring that, I'd maybe try to test it by systematizing the multiple hypothesis selection in a CV framework. Build a program to start from scratch and look at all the factors that you did in the exploratory data analysis. At the very least other seasonality periods. Make sure not to privilege the Friday->Monday conclusion in anyway. Then check to see how the cross-validated results look out-sample.

Good questions outrank easy answers. -Paul Samuelson


Total Posts: 133
Joined: Sep 2015
Posted: 2020-12-01 04:11
No, you shouldn't ignore it.

Grab yourself by your scripting-language-cumOOP-bootstraps and harvest some alpha.


Total Posts: 1
Joined: Jun 2020
Posted: 2020-12-03 14:17
To state more of the obvious (which you already know) will 'good' news come out over the weekend going forward?

How can you correct for any change in monetary policy which might be influencing your signals to date?

I'd just be worried that someone else has already seen all this and is taking me for a ride. That's my first thought.


Total Posts: 24
Joined: May 2020
Posted: 2020-12-03 14:45
Though the long term trend of the asset is increasing but I have backtested it on 15yrs of daily data and results on other configs like buy on thursday and sell on friday are very poor.
The performance is way better than the arima and other econometric models i had been slogging over past couple of months.


Total Posts: 24
Joined: May 2020
Posted: 2020-12-06 04:58
Can anyone point me on how to combine signals for a single asset?

One interesting strategy though :

This has worked better than other few things i have tried but seems counter - intuitive.However, this assumes cconstant exposure always for buy and sell. Will making exposure a function of net signal strength necessarily add value?


Total Posts: 1345
Joined: Jun 2005
Posted: 2020-12-06 10:01
Your formular does not use scaling, indeed.

Interesting idea about signal combination is here.

... What is a man
If his chief good and market of his time
Be but to sleep and feed? (c)
Previous Thread :: Next Thread 
Page 1 of 1