Nash-Sutcliffe Model Accuracy Metric

A metric to measure the ability of a model to predict observed values

Nash-Sutcliffe model accuracy statistic. Nash-Sutcliffe coefficient is an indicator of the model?s ability to predict about the 1:1 line between observed and simulated data.

The Nash-Sutcliffe coefficient is calculated as:

sum(Qobs - Qsim)^2
NS = 1 - ----------------------------
sum(Qobs - Qmean)^2

With the Nash-Sutcliffe measure, an r-square coefficient is calculated using coefficient values equal to 1 indicate a perfect fit between observed and predicted data, and r-square values equal to or less than 0 indicate that the model is predicting no better than using the average of the observed data. Thus Ockham's razor should be applied.

Dimensions of Observed data array does not have to equal simulated data. An intersection is used to pair up observed data to simulated. This function is setup to pair up the first column in the observed data matrix and the first column in the simulated data matrix. Data values are located in column 2 of both matricies.

One critical assumption to this measurement is the data are normal.

Syntax:
[NSout metric_id] = nashsutcliffe(obsDATA, simDATA)

where:
obsData = N x 2
simData = N x 2

For convenience, I?ve also posted a skillscore function which is a derivation of the Nash-Sutcliffe. Both these functions submitted are not much more than a novelty given the simplistic nature of the statistic. And my plug on a highly valuable statistic I submitted a while ago, is non-parametric trend test, the Mann-Kendall Tau-b (named ktaub.m in the Earth Sciences category). Besides the dire need for such a statistic in MATLAB, the design has some elegance in it.