They shure do seem a bit extreme. The data shows the ambient light measurement during an afternoon. By nature, the signal is noisy and can have fluctuations (in case of a shadow or cloud) faster than the nyquist frequency of my measurement. As proposed by @RoyiAvital I could have tried to fit a model. However with a model, It is difficult to regard the effects of dawn and mist.
The main purpose of filtering the data is to remove fluctuations faster than a certain frequency. Thanks to your help, I could do it. For me, it was a pleasure to learn more about digital filters!