"keep best # of results" impact on performance on optimization

[MojoJojo]

Hi Brandon,

thank you very much for your help.

https://ninjatrader.com/support/help...ionPerformance

Also, have a look at programming the strategy to be compliant with IsInstantiatedOnEachOptimizationIteration=false for an additional boost in performance.

https://ninjatrader.com/support/help...niteration.htm

I didn't know that the option was enabled by default, after the change the optimization completed in about 14 minutes. I'm very happy with the results.
My system was able to handle NinjaTrader consuming 48GB of RAM, depending on the bar data and combinations that may not always be the case though.

I'd assume that it would not be necessary to keep all backtest results in memory to get better performance, but for my needs the trade off is fine.

[NinjaTrader_BrandonH]

Hello MojoJojo,

Thanks for your post.

The higher the 'Keep best # of results' the longer it will take for an optimization to run in the Strategy Analyzer window.

There would be an additional layer of comparisons between each optimization iteration each time 'Keep best # of results' is increased, so there would be additional processing time.

Reducing the 'Keep best # of results' property will improve the optimization performance and amount of time it takes for the optimization to run.

The tips below describe what you can do about the different factors that affect optimization

https://ninjatrader.com/support/help...ionPerformance

Also, have a look at programming the strategy to be compliant with IsInstantiatedOnEachOptimizationIteration=false for an additional boost in performance.

https://ninjatrader.com/support/help...niteration.htm

[MojoJojo]

Hi,

I've run two optimizations with 50 and 1000 as "Kepp best # of results" values with intra day bars and a YTD range (88000 combinations), no other parameter changed.
The time increases from 20 minutes to 1 hour and 15 minutes with 1000 results.

There is plenty of RAM still available. I naively assume that backtesting is far more computationally expensive than filtering which results to keep based on 1 metric.

Unfortunately I couldn't find more information on the topic via the forum search.

Maybe this could be improved in a future release?