Unity's Forward Rendering Path is performant, and is currently favored for VR.
Forward rendering can be faster than deferred, it works well when you don't have an excessive number of dynamic lights, and it supports transparency for objects. It's also worth noting that forward rendering is supported on Windows, Mac, Linux, iOS, Android, and console build targets.
Also, according Alex Vlachos of Valve (GDC 2015 and 2016), Valve's teams currently favors forward rendering because it works with MSAA (multisampling anti-aliasing). Deferred rendering is not compatible with MSAA, because "lighting decisions are made after the MSAA is 'resolved' to its final image size".
Forward rendering differentiates itself from deferred rendering in the way that it treats and computes lighting. Specifically in forward rendering, a light may be rendered on a per-pixel (PP), per-vertex (PV), and/or SH (spherical harmonics) basis, which are listed in order of descending computational expense and accuracy. Because all lights are not necessarily handled on their own rendering passes, this allows for predictable rendering overhead.
Whether a light is applied PP, PV, or as SH is dependent on whether brightness, type of light (directional, point, spot), importance (the setting), and the "Pixel Light Count" in Quality Settings, as described in the documentation.
Per frame, it should be noted that lights will be categorized PP, PV, or SH, and since, be default, only the directional PP light from the base pass will cast shadows. Unfamiliarity with how settings will impact the categorization and treatment of lights may result in erratic shadow behavior, or a dimishing performance benefit over other rendering paths.
Maximizing the performance benefit of forward rendering relies on staging the scene to take full advantage of the Base Pass. The Base Pass will render objects with 1xPP, up to a max of 4xPV, and any number of SH lights. Every subsequent PP light will happen on additional passes.
Additionally, with the reliance on PV lights, 3D artists should be deliberate in both the variation and maximum size of tris in lit models.