› ›

One piece of advice for writing a trading system...

ptunneyptunney Posts: 246
in How To

If you had one piece of advice for writing a trading system what would it be?

For me, the best advice is to develop your initial idea keeping it as wide as possible to encompass as many symbols as possible.
So, if you only want to trade symbols that have gapped overnight up or down 5%, cut that to 3%.
If you only want to trade symbols that have an RVOL >3 at 10am, take those over 2 instead.

Once you get something that shows promise, break it into 4 groups. NYSE / NASDAQ LOW VOL / HIGH VOL.

Those four different groupings behave quite differently. Trying to create a model to fit both NYSE HIGH VOL and NASDAQ LOW VOL at the same time is impossible.

Do you have any top tips?

Comments

  • aj165602aj165602 Posts: 105

    Have a strategy for dealing with overfitting.

    I have a question, please: do you allow methods like bagging that use the whole dataset (and thus have a degree of forward-looking bias) or do you prefer the use of holdout data?

  • aj165602aj165602 Posts: 105

    Hi Paul,

    I’m thinking specifically of using Michaud’s resampling scheme to generate a more robust risk allocation.

    Strictly speaking, however, a trader could not decide on this allocation in real time.

    I guess what I’m trying to ask is whether any deviation away from walk-forward optimization is acceptable?

    Many thanks!

    Antony

  • spotlight1spotlight1 Posts: 2

    Antony,

    I've had success using a walkforward out of sample variation on bagging - where you'd recompute specified parameters of you strategy every month/quarter/etc. for the following month/quarter/etc., while only using the data seen before that. This way you don't have look-forward bias but you are able to optimize your model based on the most recent market data. The difficulty lies in finding these parameters and an acceptable range/learning rate to change them each time you step your analysis forward.

    I haven't had much experience with Machaud's re-sampling but after doing some research it appears that it sounds great in practice, but probably wont lead to good results when trading live due to the precision of information needed.

    Cheers,

    Trevor

Sign In or Register to comment.