Published in:Darden Business School Working Paper ; No. 3867894
Extent:
1 Online-Ressource (24 p)
Language:
English
DOI:
10.2139/ssrn.3867894
Identifier:
Origination:
Footnote:
Nach Informationen von SSRN wurde die ursprüngliche Fassung des Dokuments May 21, 2021 erstellt
Description:
Trend indicators aim to visualize the effect of input assumptions and parameters into model output. Their use is increasingly important, especially when the output is generated by a black box algorithm. We investigate the properties of several trend indicators used in simulation and machine learning, and discover that not all give consistent results when applied to well understood problems. We show that some indicators (individual conditional expectations (ICE), gradients, and partial dependence functions) are convexity and monotonicity consistent even when inputs are not independent; while some others (ALE plots, regression lines, correlation coefficients) may not be. We advocate the combined use of ICE and partial dependence plots, discuss their connection with Tornado Diagrams and Spiderplots, and propose two discrepancy indices to ease their interpretation