Comparing anecdotal actual retention to FSRS average predicted retention

That’s about average. FSRS probably just isn’t accurate enough.

Or should I just mentally deduct 5% of it to get my actual retention?

That’s very crude, idk if that’s a good idea.

Or is there a certain point where FSRS can’t predict retention accurately anymore?

In general, FSRS’s ability to accurately predict the probability of recall is worse for very high (>95%) and very low (<60%) probabilities. Around 80%-90% is where FSRS performs the best. We’re still trying to figure out how to fix that. We might change the shape of the forgetting curve again, for the third time, but there are reasons why that’s not necessarily a good idea. At least if our goal is more than just minimizing RMSE at any cost.

1 Like