Generators range from simple to elaborate: a random walk with drift, a garch process to reproduce volatility clustering, a block bootstrap of real returns, or a generative model trained on historical paths.
The most valuable use is negative. Run your strategy on 1,000 pure random walks that contain no exploitable structure. If it makes money on a meaningful share of them at your chosen parameters, your test procedure produces false positives and you have learned something important about your process.
The limitation is that synthetic data contains only the structure you put in. A generator calibrated on 2015 to 2019 will not produce a March 2020, so it cannot tell you how the strategy behaves in one.
Related: monte-carlo-simulation, block-bootstrap