5% buffer over the Q5

[casadoj]

Nina Bosshard mentioned in a previous meeting that, while sending formal notifications, they apply a buffer of 5% over the 5-year return period ($Q_5$) to avoid false alarms when the forecast is slightly over the threshold.

In the results I've seen that there are many events that are well captured by the simulations, in general terms, but they deviate slightly from the "observed" discharge, so what it could have been a hit becomes a miss or a false alarm. The 5% buffer could be applied to account for these cases in which a minor deviation between forecast / observation causes a change in the skill.

[casadoj]

Implemented.

I included a reducing_factor ($\rho$) parameter in the configuration file. This paramenter takes values from 0 to 1 and multiplies to the Q5 to create three classes, instead of the binary exceedance/non-exceedance:

• non-exeedance (0): $Q \lt \rho \cdot Q_5$
• exceedance over the reduced Q5 (1): $\rho \cdot Q_5 \le Q \le Q_5$
• exceedance over Q5 (2): $Q_5 \lt Q$ I have applied this new procedure to both the reanalysis and the forecast discharge. The 3-class datasets need to be reprocess to convert them into the necessary 2-class dataset. I have created the function recompute_exceedance to do so.