There is a differential and an integral model for W’ where the differential model is recovering much faster but has a lower max value. The Integral model will rather under-estimate W’ for repeated shorter efforts and align better with longer efforts. The differential model is opposite and will in most cases underestimate single longer efforts. There is a deeper discussion about it here:
https://markliversedge.blogspot.com/2014/07/wbal-its-implementation-and-optimisation.html
Anyway, it will never be spot-on. But it helps to design more or less accurate HIT intervals.
Once you realize the above and the fact that your FTP differs from day to day, it’s easier to understand the sometimes big errors in W’.