Well I suppose its all a matter of perspective. There are some combinations that barely produce the minimum required trades during WFO cycles. But that's not the reason I use oversampling. My feeling is that oversampling in general is a way to improve test results by reducing the effect of random entries with your logic (ie, success due to luck only).

My understanding is that oversampling will never harm the accuracy of your simulation result. On the contrary, it adds more realism to the simulation, in my opinion.

In another thread somewhere on this forum, jcl confirmed exactly how oversampling works, but I will try to explain it again here from memory:

As an example, if your strategy looks for entries on BarPeriod 15, that means during simulation, it would only look at one price sample every 15 minutes. That should theoretically be fine, as long as you have "enough" trades overall to be statistically significant when looking at the Performance Report analysis.

However, by using the NumSampleCycles oversampling feature, you have the opportunity to introduce more simulated trades "while maintaining the trend, periodic behavior, and other characteristics", per the manual.

In laymens terms, the simulation looks at the "what if" scenario of where exactly the 15 minute bar begins. If the strategy logic is truly robust, then it shouldn't really matter too much if the 15-minute bar begins at 9:15, 9:16 or 9:17. That is what oversampling does -- it shifts where that demarcation point is, at least for simulation purposes.

So my understanding is that it runs and re-runs the strategy assuming the demarcation point of the bar is at 9:15, 9:16, 9:17, etc... and then takes the average result of those tests.

Technically speaking, I believe it works like this:
NumSampleCycles=5

cycle 5 --> 9:19 bar boundary
cycle 4 --> 9:18 bar boundary
cycle 3 --> 9:17 bar boundary
cycle 2 --> 9:16 bar boundary
cycle 1 --> 9:15 bar boundary

If the strategy logic is robust enough, it shouldn't matter too much where the bar boundary is. So oversampling helps validate robust logic.