Back to Community
Insider transactions based models - OOS tearsheets

Just a thread to track OOS tearsheets of models in the Insiders transactions Challenge, so not to 'clutter' the main challenge thread too much with 'OOS' tearsheets. OOS in quotation and to be taken with some grain of salt, as it's quite possible to train on this period as well.

The first one below is of my Model 3B (non-linear interactions with 'Value factors'). OOS period: 2005 - 2014.

Loading notebook preview...
13 responses

Hey mate,

You might have just discovered the elusive "holy grail"! You're on a roll, keep up the good work!

Thanks mate! I remain sceptical, but it does look quite nice on historical data. Will attach the OOS tearsheet of my other Insiders models once I get a chance. I doubt they’ll look as impressive.

Here's the OOS tearsheet (same period as above) for my 3A insiders model. Same factors (insiders non-additive interactions with 'Value' factors) as model 3B, just different ranking scheme.

Click to load notebook preview

Model 4A OOS tearsheet (Insiders non-additive interactions with 'Growth' factors). Looks like I overfit this model a bit unfortunately. Hoping not to repeat that in my next (5th) model. I also wish I knew of a better way of assessing level/degree of noise fitting...

Click to load notebook preview

Model 4B OOS tearsheet (same factors, alternative rankings), looks overfit as well, though perhaps slightly less so?

Click to load notebook preview

Model 5A OOS tearsheet. This one suffers greatly during the GFC & Recovery periods, possibly due to the positive exposure to generic Momentum?

Click to load notebook preview

Model 5B OOS tearsheet. Same with this one, though perhaps slightly less affected by the GFC & Recovery periods?

Click to load notebook preview

Wow, very nice!

Thank you @Enigma, I try my best not to fool myself.

Here's the OOS tearsheet of my Model 2 (pure Insiders - generalized). Still a bit overfit I reckon...

Click to load notebook preview

Joakim that first one you posted is insane. This is true OOS? It wasn't trained on this period?

Thanks Viridian! None of the Insiders factors were trained on this period but some of the 'Value' factors I'm interracting with I had trained from 2010 onwards. So, anything from before 2010 would be true OOS, if you could call it that. Will the model work on future data? I have no idea, but I hope so, as that's the whole point after all. :)

Can you submit it to the contest or is the turnover too low?

Average Turnover is 5.92% but briefly dips below 5% so I would need to reduce some of the smoothing I'm doing. That shouldn't impact alpha too much, if anything it should improve 1 day IR if anything I reckon. It meets all other contest requirements though so yes, I might submit it later (I have 1 slot open).