How often do you Re-Optimize your EA

I'm stuck in a logic loop in my head, and could use some advice. I have an EA that I created and backtested on 1 and 4 hour charts. My original hypothesis was that the shorter the time frame, the more recent the backtesting limit. In other words, backtest a daily chart for 5 years and a 15 minute chart for 6 months. In my head this takes "changing market conditions" into effect, reasoning that I would need to re-optimize more often, as in every month to 3 months. But then I started down the rabbit hole in my head...am I constantly chasing "changing market conditions"? Am I forever just optimizing for last month?

How often do any of you who trade with an automated system re-optimize that system for new settings? Or to re-confirm the same settings?

And would you agree or disagree with my original premise, that if it's a shorter time frame, there's no need to go back as far on the backtest?

It's exhausting in my head sometimes...

It depends...

First of all, try to make your EA less sensitive to changing market conditions by relating thresholds to ATR, express your SL/TP as a percentage of price rather than absolute number of pips/points, lot sizes as a percentage of current equity, etc. However, that brings new challenges: should I use standard ATR(14) which changes surprisingly a lot (you try it on a chart!), or use an ATR(252) as the average movement throughout the last year but slow to react to changing conditions.

To find a good lookback period for your EA, you could do Walk Forward Optimizations with various lookback periods and see what works best for you. I have been told some people use a moving window of the last 50 trades as a lookback period, others only 2 months. Personally, I typically use a moving window of the last 4 years as a lookback period.

The re-optimization interval should be irrelevant, as the settings should reflect the optimal settings over the current lookback period. In theory you could re-optimize after every tick, or every bar. That would require an unreasonable amount of computing power, and keeps finding the same optimal settings for many iterations before enough old data has shifted out and enough new data has shifted into the moving lookback window to make another setting more optimal.

Therefore, people usually re-optimize less frequent. I know people who do this every weekend, others do it once a month, but I also know someone who does this only when performance starts to deteriorate (which is typically after a year or so with his strategy). Preferably, if you can, make the re-optimization built-in your EA and have it re-optimize itself on the fly.

Alternatively, run a WFO with a ridiculously high re-optimization frequency (e.g. hourly). That will take a lot of time but gives you the benchmark of the best possible performance of your EA. Next, try WFOs with larger re-optimization intervals. Keep increasing the interval until you notice the performance starts to deterioriate. You have now found out what re-optimization interval you should use. It is important to note that if you get wildly different results depending on the re-optimization interval, then your strategy or optimization method is probably not robust and might be curve fitted.

Most EA's with technical indicators driving trade logic can be massaged into amazingness. My version of what yoriz said: you wind up with a worthless curve fitted masterpiece that eats *fertilizer* on a walk forward test. If back test results are not fairly close across a wide range of technical settings, then it is likely not going to work in the near term.

The shorter the trading timeframe the shorter the optimization period and the re-optimization intervals. For simplification you may use general statistics: For a test result to be statistically significant it should cover at least 100 trades. You may use this number as a guideline. The length of the testing period for optimization should cover 100-200 trades and this determines your actual testing period. 6 months for M15 sounds reasonable to me.

For a 6 months test period I would re-optimize once a month. For shorter strategies (M1) I would go down to max once a week, not more often. This brings us to the next point which is more important than testing periods and re-optimization intervals.

What are you optimizing for? The answer to this question is crucial and cannot be overestimated. Your tests and optimizations are useless if the optimization criterion (aka objective function) is meaningless. Now the hard part: All test statistics provided by common retail platforms are meaningless.

Read that again, please. What does this bold statement mean? It means, it doesn't matter what you test or optimize for and how often you do it. The results provided out-of-the-box are meaningless. Why is that? Because retail platforms - and especially Metatrader 4 or 5 - are built for luring trading novices into funding accounts and hoping for the "get rich quick" miracle. There is not a single meaningful test statistic in MT4 or MT5 that you could optimize for.

Said that, what should the optimization criterion be like. First, you should always define your own criterion. MT4 offers nothing but provides an interface to user land logic, the OnTester() function. You can hook into it. MT5 provides a built-in Sharpe Ratio (calculation undocumented) which everyone knows but for optimization it's useless.

Next would be Sortino Ratio which is better but still not good enough for meaningful optimization.

I use and recommend the Calmar Ratio. There are formulas online and you can implement it into your strategy. A typical issue faced by developers is normalization to different time ranges. However, if you just want to optimze a single strategy it's not such an important issue. Just keep in mind that without normalization you cannot compare ratios of M15 strategies (intraday) with D1 strategies (daily trades).

Thanks for the suggestion. Haven't tried that one yet. I often use Martin Ratio (or the underlying Ulcer Index in some cases).

I learned the logic on only few trade signals...and then I never change the logic again. Never re-optimize it. As was said, the only change is the ATR component that specifies the volatility for risk adjustment for specific currency pair...and thats that.
Everything stays the same and one just observes the curve. With higher RR1:3 the system will encounter several consecutive DD, but that is a part of a system (several systems) portfolio. This is my take.

How many years, months is ok for backtesting to say this system is valid? - we will never know.

you need a high edge to get with 100 trades already a statistically significant calculation (you mean with statistically significant higher then 90%, or?).
i would do more 1000-10.000 trades . whatever your edge is, there is of course an optimal minimum trade amount number you need for statistically significant results.

oh, thats very easy. this number you can calculate very easy when you have a system, depends only on how much significancy you want!
but depends on your system , the kind of edge , your recalculation methods, how many people (and more important money behind) trade this edge and some general adjustments depending what backtester you use (as example metatrader you can use only with additional adjustment to make the result very exactly).