Hiya,
Alas, I'm not the best to speak on this subject (hopefully Zep or someone else can speak with more authority)
But my basic understanding is that, like most forms of data compression - it works very well with "similar blocks of same data".
So, for example, if you have a string with 50 "X" chars in it - it'll compress well, because the compression routine will store something like (use an "X" 50 times).
Whereas, if you have data that is always very different - then there's little "similarities" for it to squeeze down.
In the meantime - see here and here for some info on how the PICO-8 can compress data.
But yes, the compression a definitely dark art - even Zep said so! 😅
Hope some of this helps!