"Look Backward Re-Optimization"

Discussion in 'Strategy Building' started by traderkay, Oct 26, 2002.

  1. OK, here's the idea. You divide the data into chunks (periods). You optimize whatever parameters on the first chunk, but not trade it. Then you trade the derived parameters on the 2d chunk. Then optimize the parameters on the 2d chunk, trade 3d chunk with those parameters. I don't know if this works, probably not. There would probably have to be a high degree of stationarity in the data for this to work well. Any thoughts, experiences? Is anyone trading like this?
  2. m_c_a98


    I prefer to have no optimizable parameters.
  3. this is called "walk-forward" optimization. What else can you do, btw ? You have to optimize on history and trade on unseen data.
  4. BKuerbs



    Have a look at the works of Dennis Meyer: he wrote a series of articles about this way of walk-forward optimization in the magazine "Technical Analysis of Stocks and Commodities".

    The intervals you choose should overlap. Say you optimize on a time-span of 3 months (Jan, Feb, Mar), trade 1 month (Apr). Then pick (Feb, Mar, Apr) as your next optimization interval and trade May.

    As to the lengths and ratio of optimization and trading interval you have to test thoroughly the market you want to trade.

    The hope with this kind of testing is, that you capture a stretch of stationarity in the data, you expect the behaviour of the last opt interval to spill over into your trading intervals (That's what is called a trend in TA: up, down, or sideways).

    I did not too much testing this way, but you can see that your optimal parameters jump around a lot: a change of 100% or more is quite common.

    A word to the concept of "parameter": most people fail to understand that not only something like the number of periods you use e.g. for the RSI is a parameter. The very rules you use, breakout above whatever level, support/resistance etc. are parameters. Optimization may be done via genetic programming.

    They - like any parameter - must be optimized or they will not work. The problem is of course the same: you can only optimize with past data, the data of the future are not available.


    Bernd Kuerbs
  5. It appears to me that ideas are what works long term but the parameters change. For example a range breakout system tends to work over time but the optimal parameter will change due to the market changing. Therefore I think it is useful to optimize frequently but have a fairly long look back so the parameters only change slightly.
  6. I thought I have posted a reply to this earlier???? There is software out there that can perform this kind of tasks automatically.