Modeling optical properties of complex systems, like solvated dyes, nanostructured materials or pigment-protein complexes is highly challenging, as both coupling of the electronic excited states to nuclear vibrations and explicit interactions with the environment have to be accounted for. A successful, albeit computationally expensive approach to model linear and nonlinear spectra in the condensed phase is the cumulant approach, where the effective system-bath coupling of the electronic excitation to nuclear vibrations is computed from the fluctuations of the excitation energy in equilibrium sampled along molecular dynamics trajectories. The approach has the advantage of explicitly including the coupling of the optical excitation to its complex environment. We have recently extended the framework to approximately account for anharmonic nuclear vibrations, as well as non-adiabatic effects due to the coupling of several electronic excited states. In this talk, I will introduce the cumulant approach to computing both linear and nonlinear optical spectra of complex systems, and showcase specific applications to the modeling of absorption spectra and colors of solvated dyes and metal-organic frameworks. Finally, I will discuss how machine-learning approaches can be used to reduce the computational cost of the cumulant approach, yielding a computationally affordable framework for the calculation of optical properties in complex systems.