A single in-sample/out-of-sample split is easy to understand, but the result can depend heavily on where that one boundary happens to fall. Walk-forward validation moves the boundary forward repeatedly to test whether the same process continues to work.
Train, freeze the rules, then test the next period
For example, use six months to define the rules, test the next month out of sample, then roll the window forward and repeat. The key requirement is that the rules for each test period are frozen before its OOS result is seen.
Even re-optimization rules become part of the strategy
When to retrain, how much history to keep, and how much parameters may change all affect the outcome. Walk-forward validation is therefore not a magic button that automatically prevents overfitting.
Performance differences across market regimes become easier to see.
If a strategy holds up in some market conditions and collapses in others, the distribution of each walk-forward interval provides more information than the average return alone.