Topic: "Calibration" versus "Optimization"

Optimization vs Calibration...

I just had an "ah-ha" moment. Maybe it's obvious to most people -- I don't know.

Whenever I see people back test an EA using 2-year old, 5-year old and even 10-year old historical data I can't help but shake my head and wonder why would anyone bother testing an EA using old, stale, irrelevant data. Unless you have a time machine and plan to go back in time to trade, I don't get it.

And then something just occurred to me. I see it written all the time about "optimizing" an EA. MetaQuotes (the developers of MT4) are particularly guilty of this. Plus they perpetuate the use of the word "optimize".

Here are the dictionary's definitions for "optimize" and "calibrate":
optimize: to make as effective, or perfect as possible
calibrate: to determine or check the accuracy of an instrument

When people think they have "optimized" an EA, then it implies a process that was performed once and the EA will then perform well forever more. If you were to instead realize you have simply "calibrated" an EA, then it is understood this is a temporary check and the EA will need to be "re-calibrated" again in the future.

Perhaps because people have this misconception their EA has been "optimized" then it means back tests that date back to 2008 or 2011 or 2015 or even 2017 have relevance to how the EA will trade tomorrow. That is a totally misguided belief. And your own charts and back tests show it -- yet, most everyone chooses to ignore the obvious. Perhaps it's also due to the fact we are all lazy and really wish for it to be true that our "optimized" EA is perfect -- the alternative being we have to spend time to "re-calibrate".

Perhaps if we instead began using the word "calibrate" then it is better understood the current settings are only temporary and will need to be "re-calibrated" in the future (to take into account changing market conditions).

Does this make sense to anyone?

2 (edited by ViniQ 2018-11-11 14:44:27)

Re: "Calibration" versus "Optimization"

If you optimize it to the full range of parameters again and again, you'll have basically a different strategy (even with the same indicators) and overfit it to every out of sample segment. Instead we should just calibrate, optimize to a few parameters and limit the range to avoid such drastic change/overfitting.

This concept it's very relevant when you try out a walk forward optimization, so this make sense for me also.

Best Regards

Re: "Calibration" versus "Optimization"

This is an interesting topic. When calibrating an EA, how far back do you use historical data to complete it? 1 month, 3 month, 1 year, etc. I am finding that this is the part where I am getting stuck. I also don't agree to try to calibrate a strategy using data from 5 years ago. However, I am questioning what time period is considered 'current' in order to use the data for calibration?