Why can`t we produce better models

Pre-conference workshop « Why don’t our models work ? »
WHY CAN’T WE PRODUCE
BETTER MODELS ?
Jean-Luc BERTRAND-KRAJEWSKI, Siao SUN
OUR MODELS DON’T WORK ?
 Why are we not satisfied ?
 degree of satisfaction vs expectation
 expectation ? What is a good / satisfactory model ?
OUR MODELS DON’T WORK ?
 Modelling performance metrics
 RMSE
 NB
 Nash-Sutcliffe criterion
 modified criteria (square or root of NS)
 multi-criteria approach, e.g. (1-NS+|NB|)
 Modelling performance depends on selected criterion(a)
 Multiple-objective calibration
 Parsimonious models (number of parameters)
 AIC and BIC
ACCOUNTING FOR UNCERTAINTIES
 Uncertainty
 in measurement of both inputs and outputs
 in parameters
 in calibration
 in model structure (ignorance or uncertainty ?)
 are they separable ?
 Representativeness
 of data sets
 Variability
MODELLING
 Model = simplified representation of reality
 different from reality
 simplified
 deterministic / trends
 Reality = estimated from measurements
 part of reality (time, space)
 complex interactions of processes
 variability
 likely not strongly deterministic (level of information)
EQUIFINALITY
MODELLING
 Model = simplified representation of reality
 different from reality
 simplified
 deterministic / trends
WHY DON’T WE PRODUCE BETTER MODELS ?
 Rainfall-runoff
 Pollutants
 concentrations (pollutographs, EMCs)
 loads
 Conceptual (simple) models vs complex processes
WHY DON’T WE PRODUCE BETTER MODELS ?
 Since decades : ‘we need more data’
 Which data ? Appropriate to modelling needs ?
 Coupling measurements and modelling ?
WHY DON’T WE PRODUCE BETTER MODELS ?
 Main source : model structure uncertainty ?
 How to invent new models ?
 quality of our knowledge / understanding
 How to account for uncertainty / variability ?
 coupling deteriministic trends and processes
with unpredictable inputs ?
 grey box modelling ?