Very good point. Is there a good way to gradually decrease the impact of older data on the system during walkforward data?
The data expiring off the back end of the walkforward in sample period has just as much impact on the optimal values as the new data collected today. Wouldn't it be better give more importance and less importance to older data. Would that not reduce the system "lag" in response to changing market conditions? For example, last aug was a very "interesting" time. With a 1 year in sample period, that data would be aging off now and could result in a step change to the optimal system values. If we could somehow use a fitness function that devalued older traded compared to newer trades, it might make the system adapt more quickly and avoid rapid changes due to old data aging out of the in sample period. --- In [email protected], "Howard B" <[EMAIL PROTECTED]> wrote: > > Greetings all -- > > If the walk forward tests show poor out-of-sample results for the entire > out-of-sample period, that means the model is learning noise, not signal. > If they are good for a short period out-of-sample, then fall off, that means > the characteristics of the market being modeled are changing and you need > more frequent reoptimization -- a shorter out-of-sample time period. > > If you have a model that has in the past been profitable out-of- sample, and > it becomes unprofitable during a period of real trading, you can reoptimize > ahead of schedule. In fact, if you have valid system, you can reoptimize as > frequently as you want to. You might want to run some walk forward tests to > verify that this works for your system. If you have not tested early > reoptimizing, but do it to try to resync a system that is in a drawdown, I > recommend paper trading the new parameter set for a while to be certain that > the model just needed re-synchronizing, and was not terminally broken. > > Thanks, > Howard > > > > On Mon, Sep 8, 2008 at 2:05 PM, Mike <[EMAIL PROTECTED]> wrote: > > > Sometimes it is not so much adaptation that is needed. But rather just > > capturing the transition of cycles. > > > > Consider a house located in a region that has warm summers and cold > > winters. The variables would be climate control (e.g. air > > conditioner/heater). The variable values will change as the seasons > > unfold. To optimize comfort, we will be using increased air > > conditioning during the summers and increased heating during the > > winters. > > > > By experimenting with OOS time periods we may be able to identify > > market cycles. If you only reoptimized your values once a year (e.g. > > set for average temperature for the year), your climate control would > > likely be generally too warm in the summers and generally too cold in > > the winters. But, if you reoptimized monthly, then you would gradually > > use increasing air conditioning leading into the summer and conversely > > gradually increase heating leading into the winter. > > > > Obviously the markets are not so perfectly cyclical as the seasons. > > But, the example illustrates the point. > > > > Mike > > > > > > --- In [email protected] <amibroker% 40yahoogroups.com>, "Ara > > Kaloustian" <ara1@> wrote: > > > > > > WF testing seems like a logical step to take in creating a trading > > system. While I am not suggesting that we don't do it, I'd like to > > clear my understanding. > > > > > > It seems to me that we are looking to create a system with > > parameters that are valid "for all time". If a test is succesful in > > OOS period, then we can assume that we have found something that > > remains applicable for the future - at least in this one instance. > > > > > > Given that the markets change all the time, the obvious conclusion > > for me is that we need indicators that are adaptive... and that any > > that do not adapt will simply not work in OOS. > > > > > > The issue of providing feedback from equity curve seems valid and it > > can provide a warning about system starting to not perform as expected > > ... so regardless of type of indicators used, this kind of feedback is > > good!. > > > > > > The point I am making is that there has been very little said about > > adaptive indicators ... > > > > > > maybe that is the holy grail ... and therefore untennable ... > > > > > > > > > >
