Statistically, linearly combining independent random variables (which is what you're doing here) doesn't increase the expected value. (It *does* reduce the variance.)
In other words, if your neural networks do not have positive expectancy, then combining a large number of them doesn't help.