In the anchored version the training set grows: 2010 to 2013, then 2010 to 2014, then 2010 to 2015. In the rolling version it slides at fixed length: 2010 to 2013, 2011 to 2014, 2012 to 2015. Both then trade the following year.
Anchored windows use all available data and produce more stable parameters, which suits slow effects such as long-horizon momentum or carry. Rolling windows adapt faster and suit regimes that genuinely change, such as intraday microstructure after a tick-size change.
There is no universally right answer, so test both. If the anchored and rolling-window versions disagree wildly, that is information: your effect is not stable through time, and the live expectation should be set by the worse of the two.
Related: walk-forward-analysis, rolling-window, expanding-window, market-regime