General rule of thumb:

The higher the time-frame, the longer the back test. Depends on how long you expect to run the algo. What would be nice is to do some benchmark testing and get some standards, ie:

• >= H4 :: 10 years
• >= M15 < H4 :: 5 years
• <M15 :: 2 years

You could argue that one complete boom-bust cycle would be prudent. I think that idea is good for H1 and above, especially above H4 systems. But because technology moves on at an ever increasing pace, trading systems themselves as well as the instruments available at the beginning of the last boom-bust cycle are so different to what we have now. For the smaller time frames (particularly market micro-structure, S1 and below) you are looking at different kinds of price action dynamic shits than larger time frames).

…and ideally you want to capture huge macroeconomic events or flash crashes to stress test your system. Referendums; presidential elections; central bank announcements; interest rate changes; macroeconomic announcements and disasters – they mess up highly technical systems; especially one that run on higher time frames which is why you need to go back at least enough to capture performance during these events.

Also note that backtests don’t simulate slippage; execution errors; liquidity problems; shifts in spread; market closures and so on. A better backtest framework would incorporate simulated slippage during high volatility, changes of spread, the odd trade rejection etc etc!