I think the period of time is relative on a number of circumstances:
Are you using tick data, or just end of day?
What's the strategy? (is it using a small window or long?)
What's the Max drawdown? (ie Win:Loss Ratio)
etc.
Nobody can predict the future, but I'm a big believer in using whatever resources one can to give them an edge - that's why I backtest. The problem with most backtesting platforms is that they're insanely difficult to use - that's why I use QuantBlocks.com
Joe, as you've said, neither approach is perfect. Backtesting is faulty because past data isn't necessarily indicative of future activity. Demo trading is faulty because it's a one-way street of information instead of actual trading which is a two-way street. All of that being said, no model is...
As you've said, backtesting is inherently difficult and it's important to know what backtesting is and isn't good for. There is a new startup working on helping people backtest at Backtesting Trading Software for Quant Finance Strategies.