The data your strategy never saw while you built it, kept back to test if the edge is real.
Out of sample is the slice of history you hide from yourself while building a strategy, then use once at the end to check it. The strategy gets tuned on the in sample data. The out of sample is the last exam, data it never touched, so a good result there is a lot closer to what live trading will feel like.
Because a backtest built and judged on the same data can't really fail. You'd keep tweaking until it looked perfect on those exact trades, which proves nothing. Split your history instead, say the first 70% to build on and the last 30% to test, and you get a clean read the tuning never touched. If the edge holds on the part you never optimised, it's far more likely to be real. If it dies there, you were fitting noise. Overfitting is exactly what this catches.
Let's say you do a generic backtest with a strategy. Once you got the journal with a few hundred/thousands of trades, there's no perfect split, but 70/30 or 80/20 is common. Too little out of sample and the test is just a handful of trades, easy to pass on luck. Too much and you starved the build of data. Whatever you pick, lock it in. The moment you peek at the out of sample and go back to tweak, it stops being out of sample and quietly becomes part of the in sample. That's the one rule people break without noticing.
One split gives you exactly one out of sample test, and one test can still get lucky. Your last 30% might have been an easy market. That's why walk forward exists, it rolls the split forward again and again so the edge has to prove itself across many windows, not just one. A single out of sample pass is the floor, not the finish line.