"Martin Brown" wrote in message
On 28/05/2014 21:52, Peter T wrote:
"Martin Brown" wrote in message On

I know that's not a 'stressful' test but could you illustrate when Linest
becomes unreliable and/or returns different coeff's to a high precision
chart formula.

Add an offset of about 10000 to all the x values - this makes the
condition number of the problem much higher for a naive code. It is the
sort of thing that can happen when the x values are dates for example.

But is that a realistic scenario, 10000^5 is 10E+20 which would imply a
decimal cooefficient with at least 10 to 15 leading zeros.

When they first released XL2007 they broke the polynomial fit in charts to
make it agree with a well known PC package with the same defect!

In light testing I still can't reproduce any discrepancies.

Regards,
Peter T