Yip ...

Here is a snippet that I found useful - if you run this, you will get the same labels as the files provided.

n_base = 30

n_cell =  3

n_beam = 32

mod_cols = []

for base in range ( n_base ) :

    for cell in range ( n_cell ) :

        for beam in range ( n_beam ) :

            mod_col = f"{ base }_{ cell }_{ beam }"

            mod_cols.append ( mod_col )

Also, to pivot the sample submission (here I call it ss) I use this ... but of course, it ends up with just zeroes ...

ss [ "var_name" ], ss [ "coords" ] = zip ( * ss [ "ID" ].str.split ( "_test_" ) )

# EDIT: this!
ss [ "week"     ], \

ss [ "hour"     ], \

ss [ "base"     ], \

ss [ "cell"     ], \

ss [ "beam"     ] = zip ( * ss [ "coords" ].str.split ( "_" ) )

ss_pivot = ss.pivot ( columns = [ "base", "cell", "beam" ], index = [ "var_name", "week", "hour" ], values = "Target" )

ss_pivot.columns = mod_cols
ss_pivot = ss_pivot.reset_index ()

EDIT: I'm fighting the editing system over here ... also, important omission here that I added in.

Anyhow, this is sort of the point of my post, I was hoping to do this and end up with a bunch of explanatory variables extracted from the sample submission to use to model the target variables with, but it seems we just predict into the unseen wihtout any additional variables. So I guess, it is what it is ...

And then ... to create files similar to the input files but from the sample submission

for i in [ "traffic_DLThpVol", "traffic_DLPRB", "traffic_DLThpTime", "traffic_MR_number" ] :

    for j in [ "5w-6w", "10w-11w" ] :

        x = ss_pivot.loc [ ( ss_pivot [ "var_name" ] == i ) & ( ss_pivot [ "week" ] == j ) ]

        x [ "hour" ] = x [ "hour" ].astype ( int )

        x = x.sort_values ( "hour" ).reset_index ()

        x [ mod_cols ].to_csv ( f"../output/{ i }={ j }.csv" )

Jaw!

So nice to see you - how are you doing my good friend? Did you play any of the fossil comps? This reminds me of those. Also a bit like Sasol. Oh well ... as CB says, it is what it is.

I think, if this was a real world one, you'd struggle to beat ARIMA ... just work the stuff through that and you have decent forecasts and a relatively simple model that scales well to this sort of a thing. But ... to win a comp, you have to dig deeper I guess.

Best wishes for this one ... you are the shark of DS of course and, afaik, time series are your thing. Hope you do really well here.