Why weeks are the wrong unit
A strategy learns nothing from the passage of time; it accumulates evidence from decisions, and two strategies forward tested for the same three months can differ by two orders of magnitude in how much they have shown you.
A rule trading M15 across eight pairs might take 300 trades in three months. A daily swing rule on two pairs might take twelve. The first has a usable sample; the second has an anecdote, and both have completed a three-month forward test.
The number that matters is how many trades it takes to distinguish your expectancy from zero, which depends on the size of the edge and the spread of the R-multiples around it. A big edge with tight outcomes needs fewer; a marginal edge with a long right tail needs many more.
For typical retail strategies this lands in the hundreds, which is uncomfortable arithmetic for anyone trading a low-frequency system — and it is the arithmetic regardless of how it feels.
What does forward testing actually confirm?
It confirms three things a backtest cannot, and it is a slow and expensive way to establish anything a backtest already could.
That the implementation matches the intention. The most common finding in a first forward test is that the live system takes trades the backtest did not, or misses ones it took. That is a bug, it is found in days rather than months, and it is worth the exercise on its own.
What execution actually costs. Real spread, real slippage, real fills at real times of day. This is measurable within a few dozen trades and is the input that most often explains a gap between test and live.
That nothing peeked. A forward test runs on data that did not exist when the rules were written, which makes look-ahead bias structurally impossible rather than merely unlikely. That is a genuinely different guarantee from an out-of-sample split.
What it does NOT do efficiently is establish whether the edge exists. At ten trades a month, a 300-trade confirmation takes two and a half years — which is why out-of-sample backtesting does that job and forward testing checks the things backtesting cannot.
Paper or live money?
Paper trading tests the rules; a small live account tests the rules, the execution and the trader, and the third of those is the one that most often diverges.
Demo fills are typically better than live ones — no slippage pressure, no requotes, and an execution model that is essentially the broker being polite. A strategy that works on demo and fails live is frequently discovering the difference.
The psychological difference is larger still. The behaviours that cause traders to deviate from their rules — skipping a setup after three losses, moving a stop, taking an unplanned trade — do not appear when nothing is at stake, and they are the mechanism by which a good strategy produces bad results.
The common compromise is a defensible one: paper trade until the implementation is verified and the trade log reconciles against the rules, then go live at a size small enough that the education costs less than it teaches.
How do you know when to stop testing?
Decide the stopping criteria before you start: the trade count you will accumulate and the result you will require, both written down in advance.
The reason to write them first is that they are unwriteable afterwards. A forward test with no predefined endpoint runs until the results look convincing, which is a procedure guaranteed to produce a convincing result eventually regardless of the strategy.
A workable form is a pair of thresholds: a minimum number of trades, and a floor for expectancy below which the strategy is rejected rather than adjusted. The rejection branch is the important one and the one usually missing.
Then hold to it. Extending a forward test because the results are close, or cutting it short because they look good, both convert a measurement back into a search — and a search you conducted on yourself is the hardest kind to notice.