Also, I haven't gone through the first paper, but that second link just seems to show how they generally conform to IEEE 754, except for three cases (div/0 errors instead of infinity, NaN errors immediately instead of allowing further computations, and [this one is a bit more important] they don't implement denormalized numbers). I'm not sure why that would mean Excel is terrible at floating point math.

This has always been one of my concerns when Excel is used much beyond 'electronic squared paper'. The interwoven data and code nature of the tool makes testing and verification more difficult - certainly not impossible but going against the grain of the tool. Excel is used by a diverse range of people, many of whom have limited software skills and could do with a tool that enforced the right behaviour.

Also I think this shows the need for openness when people make such important claims based on data. Particularly in this case where all the data and calculations could easily be zipped up and put on a website.

It is so easy to make a mistake on a big spreadsheet.

A much better approach would be to write user defined functions.

Here's a tip: if your results are reproducible only in Excel, something is terribly wrong.