The procedure: take 2010 to 2013 as the fitting window, choose the best parameters, apply them unchanged to 2014, and record those trades. Then fit on 2011 to 2014, trade 2015, record. Repeat to the end of the data. The concatenated trading windows form a track record in which no parameter ever saw its own future.
That record answers a more useful question than a single out-of-sample test: not is this parameter good, but does my whole process of choosing parameters produce something that works. It also produces a walk-forward efficiency ratio, the out-of-sample return divided by the in-sample return over the same windows; anything under about 0.4 suggests the fitting is mostly chasing noise.
The cost is honesty about compute and sample. Ten windows of one year each is ten out-of-sample years, but the parameters change every year, so the effective sample per parameter set is small. Use it alongside parameter-sensitivity rather than instead of it.
Related: anchored-walk-forward, rolling-window, out-of-sample, parameter-sensitivity