This paper develops smaller statistical models from a finite-state Markov representation of a dynamical system. A modified community-detection method groups states to isolate the dynamics most relevant to a chosen timescale. A complementary construction estimates a continuous-time generator that represents transitions across a wider range of timescales. The methods are demonstrated on stochastic and chaotic examples, the Kuramoto-Sivashinsky equations, and experimental fluid-flow measurements. Readers get tools for choosing the complexity of a reduced model and evaluating whether it preserves the temporal statistics that matter for their application.