There are a number of reasons, as @tardis-maker suggested in the other answer.
Here are a few more:
- Grading. As stated, it can be beneficial to be looking at a display referred view of your final output that covers any number of stops above middle range, properly mapped to the display referred viewing transform
- Proper desaturation. Rendering in a scene referred model extends primaries off to infinity. This means that they do not desaturate nor reach the display referred white in any way that is familiar nor correct. Consider the case of RGB 0.001 0.001 and 3.6. In this case, the only means to have blue blow out to diffuse white is via a forced intensity shift of 1000x to bring the other primaries up past the display referred ceiling of 1.0. Consider the following demonstration of primaries, that exacerbates the issue:
In both instances, the render of intensities is identical, but only in the latter does it resemble a "photorealistic" result because of the response of the display referred view transform. Notice how the pure primaries never desaturate / converge toward the hull established by the display referred primaries limitation, and instead carry on extending toward infinity, at their identical primaries chromaticity. While the above images show the flaws of a strictly 1D viewing transform, every colour is mangled in precisely the same way, to lessening degrees as a colour's luminance draws near achromatic. If you examine the achromatic cubes, it is clear how adjacent colours should converge to display referred white.
Related to the above point, the default sRGB output view transform captures a mere two and a bit stops of light above middle grey from a Cycles render. This is entirely unnatural when compared to our learned response of examining photographic-like reproductions, which maps anywhere from six or more stops of light above middle grey to the display / output referred transform. The above images encode approximately six and a half stops of light to the display referred view, and as such, the scene properly reflects the lighting with respect to influence of bounces and other radiometric responses without cheats or hacks to bring the result into the display referred domain.
Congruent with what @tardis-maker states above, all View transforms operate strictly on the one-way output of the viewing transform. That is, it is a means of viewing a baseline look without adversely impacting your scene referred internal values. It is entirely non-destructive in terms of evaluating an image as compared against a node chain, which corrupts and mangles values as the data flows along the chain.
It should be noted that the film emulations in Blender, although based off of solid data, are not implemented in a manner that makes them useful in the terms above. The breadth of why is beyond this post. Instead, they should be considered as nothing more than Instagram filters as a random creative dice roll.
http://www.oscars.org/aces/spotlight-aces-vfx Pay attention to where he discusses cheating values.