Measures of Model Performance Based On the Log Accuracy Ratio
Measures of Model Performance Based On the Log Accuracy Ratio
复制标题
DOI：
10.1002/2017sw001669
复制
发表时间：
2018-01-01
影响因子：
3.7
通讯作者：
Welling, D. T.
中科院分区：
文献类型：
--
作者：
Morley, S. K.;Brito, T. V.;Welling, D. T.
连续量的建模和预测的定量评估使用多种方法。我们回顾现有的文献描述的预测准确性和偏差的指标，集中在那些基于相对误差和百分比误差。在这些精度指标中，平均绝对百分比误差（MAPE）是许多领域中最常见的指标之一，在最近的空间科学文献中得到了广泛应用，我们重点介绍了MAPE的优点和缺点以及提出的替代方案。然后，我们引入了对数准确率，并从中推导出两个度量：中位数对称准确度和对称有符号百分比偏差。还提出了使用对数精度比估计乘性线性模型的扩展的鲁棒方法。所开发的指标被证明是易于解释的，强大的，并减轻其更广泛使用的同行的相对误差和百分比误差的基础上的主要缺点。它们的使用说明与辐射带电子通量建模的例子。
Quantitative assessment of modeling and forecasting of continuous quantities uses a variety of approaches. We review existing literature describing metrics for forecast accuracy and bias, concentrating on those based on relative errors and percentage errors. Of these accuracy metrics, the mean absolute percentage error (MAPE) is one of the most common across many fields and has been widely applied in recent space science literature and we highlight the benefits and drawbacks of MAPE and proposed alternatives. We then introduce the log accuracy ratio and derive from it two metrics: the median symmetric accuracy and the symmetric signed percentage bias. Robust methods for estimating the spread of a multiplicative linear model using the log accuracy ratio are also presented. The developed metrics are shown to be easy to interpret, robust, and to mitigate the key drawbacks of their more widely used counterparts based on relative errors and percentage errors. Their use is illustrated with radiation belt electron flux modeling examples.