Tick interval selection and optimization

Discussion in 'Automated Trading' started by JangoFolly, Sep 4, 2005.

  1. I have an automated intraday strategy that I'm currently backtesting on 9 months of level-one tick data for the mini-Dow. It’s a triple timeframe (or, in this case, tickframe) approach à la Dr. Elder with x, y, and z ticks per period, and I’m using canned indicators (MACD, Williams %R, EMAs) with non-standard parameters. This system is very similar to my normal discretionary approach for trading the YM, which has been moderately successful to date. I have done some tweaking of the indicator and parameter combinations which constitute an entry setup based on my testing, but I’m not too far from where I started and I’ve tried to keep the tweaks as general as possible to avoid over-fitting my model to my test data. I am, however, a little concerned because my results change dramatically for the worse when I change the values of the three tickframe intervals. As I mentioned before, these tickframe values are based on a successful discretionary approach rather than having cycled through a range of values looking for the most profitable tick interval. On the other hand, I find it somewhat unlikely that my discretionary system just happened to utilize near-optimal tickframe values. I have also run automated system over a similar range of data for the mini-S&P, 10-year T-note, and Bund futures. None of the three resulted in a profitable system, but I’m less worried by that since each market has its own rhythm.

    I do plan to get some additional YM data and test the model over a new range of data, as well as begin testing on live data in real time. I would appreciate any thoughts you have on this matter.

    Thank you.


    Regards,
     
  2. Perhaps you should use some sort of continuous time model.. e.g., the time between events is another parameter. quantatizing tick data is like putting a square peg into a round hole.

     
  3. fader

    fader

    (1) "As I mentioned before, these tickframe values are based on a successful discretionary approach rather than having cycled through a range of values looking for the most profitable tick interval." - i see your point, but i would also guess that in your discretionary approach you must have tried a few different tick intervals before it worked... it's unlikely that you have just picked them in your discretionary and they worked immediately.. - so still you may have "fitted" these timeframes in your discretionary approach although you did not go through the more rigorous testing that you are doing now.

    (2) 9 months of data is not sufficient in my opinion - especially because you can't use it for forward testing since your parameters seem to be fairly locked - i backtested a lot using 4 years worth of data and found that most if not all strategies' performance goes through cycles (which correspond to the cycles in market behavior) - the cycle length can be a few months at least or much longer - i.e. your current backtesting period may be covering a market cycle which is particularly favorable to the behavior your strategy is trying to capture (fyi i did not backtest with tick data, only intraday minute data)

    (3) actually i may overlook the comments above if you told me you have traded this approach discretionarily for 10 years (or X many years) and this system is a very close replica of it... - i guess in this case you are just mechanizing something that you know has worked consistently for a long time.. - that's a different matter.

    (4) on the other hand, i can agree that the short-term behavior of the other markets you have tested, i.e. ES,bonds,bund etc. is distinctly different from YM - these other markets are awash in liquidity and there are not as many deviations from the equilibrium levels - i can suggest you look at a comparable instrument, like the DAX futures for example or ER2 and see the performance might be more similar.

    (5) whatever you do, you must have a very strict plan for stopping/re-evaluating/re-working the system if it starts performing below your backtested results.

    all the best.