Maybe it is actually the PY numbers that are wrong and not the measurement system ratings ?

For some reason people always believe that RAW statistical results are more in line with reality then mathematical models smoothing the raw statistical data into properly balanced tendencies.

This is a very grave error. An understanble error maybe but a grave error nonetheless.

Compare it to this. Take you GPS track log and remove any smoothing (averaging) of the speeds, you are very likely to see more then 1 or 2 spikes in your speed data. There is simply no way you can accellerate from 10 knots to say 25+knots and back again in 10 seconds. Any real life system will be disturbed by such factors, many systems will be HIGHLY disturbed by such factors all the time, in that case we say the data is disturbed by noise. Compare the last situation with old AM radio's and walky-talky's; the cracking and hissing sounds making speach almost indelligiable. Modern devices you a score of mathematical smoothing and post processing technics to filter out the speach from the background noise/disturbance. And the measurement based ratings system actually so a very similar thing.

Measurement based rating systems force a smooth function (or graph) to be fitted to heavily disturbed measured data. This graph is then the best fitting for that dataset given its enforced limitations on smoothness. So yes, it may distort the pitch opf the "speach" at little but it is still a [censored] load better then the old way of trying to hear what somebody is using over the screams and clicks of the static (background noise).

This graph below gives such a example of fitting a very simple mathetical model to a very disturbed set of measured data.


Raw data and the real underlaying system




Smoothed data and the real underlaying system





Now any rating system that is entirely based on raw statistic data will have to work with a cloud of distributed datapoints where no simple assumptions with respect to averaging can be made.

An example : If all data is taken from fleets that have active group of skilled racers then simply averaging the raw data will produce smoothing. But when some fleets are raced much more actively (F18, A-cats) then others (Prindle 16, supercat 17) then simply averaging the raw data will results in relatively large biases with respect to the real (yet unknown) underlaying system (situation).

Of course, in our field of interest, sailboat races, we can simply not assume that the skill level of racing crews is the same over all racing fleets (boat types). Actually we know for certain that very large differences exist EVEN inside a single class ! Example; Mischa Heemskerk coming over to the USA grabbing a crew from the beach and putting in a line of straight bullets in any US F18 regatta while he can't be that chavalier about it in any EU F18 race.

Now with sailwave we are proposing to mix this EU data with US data and expecting improved accuracy ?

I will tell you what will happen, you will loose accuracy as you will be averaging the relative difference in fleet skill between EU and USA and deduct that from the REAL F18 designed in performance which will at least be the same or just above the EU level.

Having more data simply does not garantee higher accuracy. There is such a thing as poor and good data.

Next points made in my follow-up posts.

This "system identification" stuff is actually a standard skill in my line of work.

Wouter