Walk Forward Optimization

Discussion in 'Strategy Building' started by Splat, May 5, 2003.

  1. Splat


    What is a reasonable period for a system be reoptimized?
    1 week? 2 weeks? 3 weeks? 1 month? 3 months? 6 months?

    Why am I asking this question? Well the market behaviour
    and general sentiment seems to change continuously every few
    days , trends seem only to last a few days. So
    I was thinking that maybe a reoptimization could be performed
    *once* a week. Would this seem reasonable? Please give
    me your opinions.


  2. nitro


    When I did programming for a hedge fund, it would be done every 6 - 8 months. That was back in '89 -90 though.

  3. Splat


    Wasnt the market in an uptren at that time? When the market
    is in a trend I can well understand doing optimizations
    every few months but in an environment such as we have
    today, surely we need to do reoptimize more often?

  4. You want your reoptimization period adjusted for the timeframe you are trading. It would be silly for a swing trader taking 3-5 trades a week to reoptimize weekly.

    If it were me I would use a reoptimization period that typically covers a statistically significant number of trades or setups. For me, with 15-20 trades a week nowadays, I would probably use at least two weeks. Back when I was taking longer-term position trades I'd use 2-3 months as the walk-forward.

    A slightly more scientific method might be to determine if there's any cycle to your 'system', especially in it's drawdown, or in the market characteristic you're trying to take advantage of. You could then reopt on a half-cycle time frame (similar to classic Nyquist sampling theory).

    However, if the success of your system depends greatly on the optimization cycle you may have other issues in the future. Systems that need constant tuning to maintain performance often blow up. Be prepared.
  5. Splat


    Thnx for your reply. That was the kind of thing I was thinking
    already. I have written a system to initially trade on the dow
    which generates 5-6 swing trades per month. The win ratio
    is at least between 60-100%. The system finds the possible
    trades which have the biggest momentum.
    Currently I am finding that it needs an optimization at least
    once a month and works very well.
    I am planning on generating the trades with tradestation and
    executing the trades with a UK spreadbetting company.


  6. Splat


    Forgot to add that I am using a 30min timeframe.
  7. hklein


    Is there a software that has it included. I would like to test many markets and then trade them. Would the performance be better with it? Any help welcomed.
  8. nitro


    I agree. Just be aware that the more you optimize, the more you curve fit.

  9. Splat


    Have a look at the search faciity here. There is a lot of
    info on optimization.


  10. man


    best thing IMO is to embed reoptimisation into the strategy as such. then you do not have to worry since rebalancing is part of the plan. One way to do it is to do the rebalancing at fixed intervals of time. another could be to make it event dependent. you could define events that should trigger optimisation, like unusual market behavior as changes in volatility, daily range, trendiness or you could use your equity curve to indicate when reoptimisation should take place.

    the problem is definitely again the temptation to overfit. make sure that you have enough optimisation points if you use event driven optimisation, otherwise you are the perfect "fitter".

    #10     May 6, 2003