A forward test is the opposite of a backtest. The strategy is fixed, a start date is set, and the strategy is run on prices as they arrive, on a demo or a small live account, for months. Nothing about the strategy is changed during the test. The results are whatever the market delivered to a system that could not have been fitted to it, because the data did not exist when the system was written.
Why it is the fair test
Every limitation of a backtest disappears. The spread is the real spread, widening included. The fills are real fills. Commission and swap are charged. The parameters cannot have been chosen on the data, because the data came afterwards. And the strategy tested is the one that was committed to on the start date, not the best of many tried. A forward test is a backtest that cannot be faked.
Its cost is time. A backtest over ten years takes minutes. A forward test over ten months takes ten months, and ten months is the minimum for it to contain more than one kind of market.
How to run one
Fix the strategy and its settings and save them. Choose a demo account at the broker to be used live, or a small live account, and record the start date and balance. Decide in advance what would count as a pass and a fail: a return above some level with a drawdown below some other, over a stated period, or a result consistent with the backtest. Then leave it alone. Changing settings mid-test restarts the test.
If the account is connected to a third-party verification service from the start date, the test is public and the record cannot be edited, which is the strongest form the evidence can take.
A developer finishes a strategy on 1 March, saves its settings, opens five demo accounts at one broker, one for each risk setting, connects them to a verification service and starts them the same day. On 1 September the accounts show six months of trades, two of them with drawdowns of 9 and 13 per cent at the medium setting and a return in line with what the backtest implied, at costs the backtest could not model. Nothing was changed in the six months. That record is worth more than any backtest of any length, and it is still only six months.
An EA vendor's forward test, published from a sealed start date on a verified account, is the evidence to look for. Check that the start date is stated, that the settings were fixed, that the record runs continuously without resets, and that the vendor set pass and fail criteria before the test rather than describing the result afterwards.
When testing an EA yourself, do the same. Fix the settings, note the date, use the broker you will use live, and decide before you start what result would make you stop. Then let it run through at least one period you would rather it hadn't.