{"id":"W3188169536","doi":"10.1029/2020wr029001","title":"The Abuse of Popular Performance Metrics in Hydrologic Modeling","year":2021,"lang":"en","type":"article","venue":"Water Resources Research","topic":"Hydrology and Watershed Management Studies","field":"Environmental Science","cited_by":298,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Saskatchewan","funders":"Global Water Futures; University of Saskatchewan","keywords":"Hydrological modelling; Environmental science; Hydrology (agriculture); Computer science; Geology; Geotechnical engineering; Climatology","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.1854398,0.001343035,0.002462732,0.004675631,0.002714756,0.01210021,0.006572214,0.005934151,0.001085771],"category_scores_gemma":[0.4512995,0.0005953625,0.002055142,0.006811505,0.02005353,0.01221059,0.004408857,0.01206029,0.0003930439],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01088054,"about_ca_system_score_gemma":0.008458135,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01000906,"about_ca_topic_score_gemma":0.006723693,"domain_scores_codex":[0.7382663,0.1960561,0.0154481,0.009029524,0.03868967,0.002510458],"domain_scores_gemma":[0.3312594,0.5543321,0.02149663,0.0221254,0.06941952,0.00136691],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0001473128,0.00004596152,0.003652205,0.0008365283,0.0002354495,0.0001415709,0.002291754,0.02416901,0.0001748729,0.8637649,0.04512496,0.05941543],"study_design_scores_gemma":[0.00004706502,0.0003075661,0.005396303,0.003763094,0.0001467051,0.0003261739,0.002605448,0.08742171,0.00209949,0.7495428,0.148024,0.000319748],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.03190501,0.07105554,0.4790589,0.3607392,0.01753635,0.0002452499,0.0007997779,0.0006719421,0.03798807],"genre_scores_gemma":[0.7521288,0.02543495,0.1199543,0.06281754,0.03352859,0.0006081129,0.0003029313,0.001072445,0.00415233],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8145602,"threshold_uncertainty_score":0.980711,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.05464573691619524,"score_gpt":0.2947246862389659,"score_spread":0.2400789493227707,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}