Kenneth, thanks for your continued explorations. I've been looking at this most of the day, and I think I finally made a decision.

Looking in detail at the differences between the original data (red), my original algorithm (blue); graff's agorithm (black) and your latest algorithm (green); there is one place where I think your algorithm is clearly superior: the bottom left corner of the top right zoom, where your algorithm splits the difference between mine and graff's; and I think is clearly a better reflection of the underlying trend of the original data.

However, as you pointed out, it does 'invent' points, which neither of the other two do. I also agree that in most other cases yours has a tendency to bias high.

So after much umming and aahing, I'm going with graffs algorithm. Thanks.