I know the "Kalman filter" mostly in its application as an "observer" in control systems. However, its use there is indeed to estimate the values of certain states that can not be directly measured. It is as such an estimation technique.

After thinking about it some more, it does strike me as a parameter estimation (weighting coefficients) procedure after the designer has chosen the model structure. As such it falls in the very same (regression) approach as Texel and SCHRS deploy, although they use a different regression technique.


I just looked it up in Wikipedia and it reads like something I should understand. Sadly I think some pages in Wikipedia have been hijacked by mathematic nerds and rewritten in their own special secret language.

Leave it to formally trained mathematicians to write down simple stuff in the most confusing way.

Anyway, the general approach you are suggesting is in my opinion the best way to build a rating system. Choose a model structure of a relative low order and then use an estimation technique (like Kalman filtering) to estimate the weights. Of course all ratings will then have to be calculated in exactly the same way as Texel and SCHRS do, by measuring boats and punching in the specs in the regressed formula.

This is significantly different from how some posters understand this intention in sailwave.

Summerized it is what Texel and SCHRS are doing only you (Sailwave) is proposing to use an estimation technique that continiously updates itself with each new data coming in.

It that is the case then you still have the problems associated with undependable datasets. What if I enter the results of my last (drunken) sunday fun club race ? It appears to me that Kalman filtering will respond to that when in reality it should not.

Interesting problem. But there are always next years !

Wouter