Item: Understanding the "True" Predictive Power of Statistical Avalanche Forecasting
-
-
Title: Understanding the "True" Predictive Power of Statistical Avalanche Forecasting
Proceedings: International Snow Science Workshop Grenoble – Chamonix Mont-Blanc - October 07-11, 2013
Authors:
- Jordy Hendrikx [ Snow and Avalanche Laboratory, Department of Earth Sciences, Montana State University, Bozeman, MT, USA ]
- Matt Murphy [ Alaska Department of Transportation and Public Facilities, Anchorage, AK, USA ]
Date: 2013-10-07
Abstract: Statistical avalanche forecasting tools using meteorological and snowpack predictors have been developed in many different snow avalanche environments around the World. The statistical approaches employed have also varied widely, e.g. correlation analysis; multivaritate discriminant function analysis; nearest neighbors; expert systems; classification and regression trees; and crossvalidated classification trees. All of these approaches require a historical data set of meteorological, snowpack and avalanche observations with sufficient unique avalanche events, which are then often grouped into periods of avalanche days. Classification trees in particular have seen much attention for both exploratory and predictive purposes in both operational and research settings. However, when used in a predictive mode outside of the original training period, it still remains unclear about the "true" predictive power of this statistical approach. The cross validated classification trees purport to provide a statistical measure of the predictive ability; but how well do these actually work outside of their training data set? This paper will explore the concepts of learning time and parameter space with specific reference to classification trees for avalanche forecasting. The predicted statistical accuracy using the 10-fold cross validation technique has been compared to the "true" accuracy when tested outside of the training period. These concepts have been examined using a 30 year data set of avalanches and their associated meteorological observations by varying the training and the testing data set lengths and composition. This analysis indicates that the training period is critical for robust predictive results. While temporal length of the training data is important, parameter space coverage is likely more critical for robust predictions outside of the original training period.
Object ID: ISSW13_paper_O4-09.pdf
Language of Article: English
Presenter(s): Unknown
Keywords: avalanche forecasting, statistical forecasting, classification trees
Page Number(s): 191-194
Subjects: avalanche forecasting avalanche prediction classification trees
-