Well that depends. On backtest you can reliably test for tens of years of data and that is what's really important. On live testing you would have to keep your EA active for literally tens of years to get the same testing results. Therefore that is not an option.

I still think that when using correctly, ie when you know what to do and what not to do (test scalping for example) backtesting results are good. If you are testing for example strategy that keeps the order open for at least couple of hours then the random fluctuations at the price at open or close don't really matter because more the trades, more the results will be similar to the tester.

Live trading for couple of months proves nothing because markets may be favourable to the hardcoded things on your advisor, not that your advisor works on the long run.

I think the main problem is that people curve fit and then wonder why testinfg results don't reflect the real world results. That's why you have to test your strategy on multiple currencies. If it works on more non interrelated pairs than it doesn't then you have a good strategy. If it only works on the currency you have been developing the system with, you have a curve fitted system.

Simple. And hard at the same time.