I found

this blog post where the formula is: CL=(C1*C2)/(C1+C2)+Cs. If the crystal has a load capacitance of 20pF the capacitors would be just about 35pF if the stray capacitance is guesstimated to 2.5pF, because 20=(35*35)/(35+35)+2.5.

But when I look into the

spec sheet for an ATmega328P (section 13.3), the capacitors is to be calculated with the formula: Ce+Ci=2Cl-Cs where Ci is the internal capacitance of 18pF on XTAL1 and 8pF on XTAL2. With the same crystal and stray capacitance as before that would result in two different caps of each ~20pF (XTAL1: 20+18~=20+20-2.5) and 30pF (XTAL2: 30+8~=20+20-2.5).

The difference between the two formulas is not insignificant and I'm assuming that I should trust the manual. But would the first method yield better accuracy? And how would a "too large" or "too small" capacitor influence the crystal?

And finally: I'm using a 32kHz (20pF Cl) crystal with Timer2 on an ATmega328P and it is not 100% accurate but the frequency is consistent compared to the WDT which is definately not reliable. If I remove the 22pF caps the crystal (or rather the entire code execution) becomes defunct (spec sheet dictates that Cl>6pF requires external caps).

Thanks in advance!