>Jos,
>
>The actual root problem greatly helps! If I were approaching this "noisy data" problem, I would not alter the raw data because you have no auditability. What I'd do is store all of the raw data, gather the evidence of errors and use that to get the 3rd party to provide better data.
>
>At display time I think you can choose to display "filtered results", something like ignore data points that have a deviation of N sigma from the average over a sliding window of days. What this would do is smooth the curve it would also time skew it so the plots would not show large changes in value until a day or two later. This may or may not be acceptable. You have to make sure that everyone understands the effects of the data display filters going into it.

Also, I don't see any need to do it in ten steps. Just take some honest approximation between the last value and zero, maybe half of last time's sigma. The task description says only that we may need to come closer to zero if it repeats - well, if it does, last value is already closer, and it has introduced some scatter (by being half a sigma below the previous value), so sigma may even increase, so it will approach zero even faster if this repeats. IOW, if the zero is a bona fide value, then it would be reached fast enough; if it's a glitch, it will be sigma/2 of an error or somewhere around that. And there's no limitation on the number of times needed to recognize a zero as a true value - it may become close enough fast enough.

Running this algorithm (and varying between my proposed half sigma and other possible values) against old data may show what's most fitting for the purpose.