However, I’m surprised no one has yet caught the mistake (probably just a typo) in the formula for cosine similarity in the edge weights worksheet. The formula should be :

sumproduct(…) / ( sqrt(…) * sqrt(…) )

So there is an extraneous left parenthesis, and + should be replaced with *

In mathematical terms, the cosine similarity of two vectors u and v is :

dot_product(u,v)/(length(u)*length(v))

Obviously, cos_similarity(u,v) = cos_similarity(v,u)

which means the similarity matrix in the edge weights worksheet should be symmetric.

I know this conversation is a year old, but I am still interested so perhaps others are as well. I was also looking for a solution to this problem concretely related to forecasting for hotels. Hotels have several “seasonalities” including day of week patterns, as well as a monthly type of seasonality. Hotels need daily forecasts to manage properly.

I found this recent paper on extentions of exponential smoothing. http://users.ox.ac.uk/~mast0315/TripleSeasonalEJOR.pdf . It would be great to have it “translated”…

]]>