Pure Analog: when to use days without precipitation

[jhamman]

@gutmann and I just had a conversation about the use of (non-)precipitation days in the pure analog case in GARD.

Some things to look into:

• Normalization - should non-precip days be used in mean/std?
• analog selection - should non-precip days be used in analog selections? With/without sample_analog?

[andywood]

Joe, * I think the main benefit of standardizing without zero days is that the resulting distribution is a bit more normal. * on the selection problem, I would hope the selection could determine when picking a zero-pcp day is appropriate. One of the nice properties of analog downscaling for precip is getting the clear days to be absolutely clear, rather than regressed into the neighborhood of clear. My sense is to keep them, as they may be reasonable analogs for low pcp/moisture days.

…

-Andy

On Mon, Nov 20, 2017 at 3:59 PM, Joe Hamman ***@***.***> wrote: @gutmann <https://github.com/gutmann> and I just had a conversation about the use of (non-)precipitation days in the pure analog case in GARD. Somethings to look into: - Normalization - should non-precip days be used in mean/std? - analog selection - should non-precip days be used in analog selections? With/without sample_analog? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#64>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AFCgRSrvfpc-c93P2fbR5OLjMzKwrwkLks5s4gQ5gaJpZM4QlB5Y> .

[gutmann]

On selection: I agree, that zero-precip days should be left in when selecting a collection of analog days. If a single analog day is being sampled from that collection, then zero-precip days should be left in. However, if you want to output a probability of precipitation along with some mean and error term, then it makes sense to calculate the probability of precipitation across all analogs, but the mean and residual "error" (i.e. std.dev.) should just be computed for days that have precip. Then the SCRF post-processing uses the probability to select between zero or non-zero precip, but if precip occurs, then the amount is computed based on the mean and std.dev of non-zero precip analog days.