• Documentation
    • About ​ValidMind
    • Get Started
    • Guides
    • Support
    • Releases

    • Python Library
    • ValidMind Library

    • ValidMind Academy
    • Training Courses
  • Documentation
    • About ​ValidMind
    • Get Started
    • Guides
    • Support
    • Releases

    • Python Library
    • ValidMind Library

    • ValidMind Academy
    • Training Courses
  • Log In
    • Public Internet
    • ValidMind Platform · US1
    • ValidMind Platform · CA1

    • Private Link
    • Virtual Private ValidMind (VPV)

    • Which login should I use?
  1. Test descriptions
  2. Ongoing Monitoring
  3. ClassificationAccuracyDrift
  • ValidMind Library
  • Supported models

  • Quickstart
  • Quickstart for model documentation
  • Quickstart for model validation
  • Install and initialize ValidMind Library
  • Store model credentials in .env files

  • Model Development
  • 1 — Set up ValidMind Library
  • 2 — Start model development process
  • 3 — Integrate custom tests
  • 4 — Finalize testing & documentation

  • Model Validation
  • 1 — Set up ValidMind Library for validation
  • 2 — Start model validation process
  • 3 — Developing a challenger model
  • 4 — Finalize validation & reporting

  • Model Testing
  • Run tests & test suites
    • Add context to LLM-generated test descriptions
    • Intro to Assign Scores
    • Configure dataset features
    • Document multiple results for the same test
    • Explore test suites
    • Explore tests
    • Dataset Column Filters when Running Tests
    • Load dataset predictions
    • Log metrics over time
    • Run individual documentation sections
    • Run documentation tests with custom configurations
    • Run tests with multiple datasets
    • Intro to Unit Metrics
    • Understand and utilize RawData in ValidMind tests
    • Introduction to ValidMind Dataset and Model Objects
    • Run Tests
      • Run dataset based tests
      • Run comparison tests
  • Test descriptions
    • Data Validation
      • ACFandPACFPlot
      • ADF
      • AutoAR
      • AutoMA
      • AutoStationarity
      • BivariateScatterPlots
      • BoxPierce
      • ChiSquaredFeaturesTable
      • ClassImbalance
      • DatasetDescription
      • DatasetSplit
      • DescriptiveStatistics
      • DickeyFullerGLS
      • Duplicates
      • EngleGrangerCoint
      • FeatureTargetCorrelationPlot
      • HighCardinality
      • HighPearsonCorrelation
      • IQROutliersBarPlot
      • IQROutliersTable
      • IsolationForestOutliers
      • JarqueBera
      • KPSS
      • LaggedCorrelationHeatmap
      • LJungBox
      • MissingValues
      • MissingValuesBarPlot
      • MutualInformation
      • PearsonCorrelationMatrix
      • PhillipsPerronArch
      • ProtectedClassesCombination
      • ProtectedClassesDescription
      • ProtectedClassesDisparity
      • ProtectedClassesThresholdOptimizer
      • RollingStatsPlot
      • RunsTest
      • ScatterPlot
      • ScoreBandDefaultRates
      • SeasonalDecompose
      • ShapiroWilk
      • Skewness
      • SpreadPlot
      • TabularCategoricalBarPlots
      • TabularDateTimeHistograms
      • TabularDescriptionTables
      • TabularNumericalHistograms
      • TargetRateBarPlots
      • TimeSeriesDescription
      • TimeSeriesDescriptiveStatistics
      • TimeSeriesFrequency
      • TimeSeriesHistogram
      • TimeSeriesLinePlot
      • TimeSeriesMissingValues
      • TimeSeriesOutliers
      • TooManyZeroValues
      • UniqueRows
      • WOEBinPlots
      • WOEBinTable
      • ZivotAndrewsArch
      • Nlp
        • CommonWords
        • Hashtags
        • LanguageDetection
        • Mentions
        • PolarityAndSubjectivity
        • Punctuations
        • Sentiment
        • StopWords
        • TextDescription
        • Toxicity
    • Model Validation
      • BertScore
      • BleuScore
      • ClusterSizeDistribution
      • ContextualRecall
      • FeaturesAUC
      • MeteorScore
      • ModelMetadata
      • ModelPredictionResiduals
      • RegardScore
      • RegressionResidualsPlot
      • RougeScore
      • TimeSeriesPredictionsPlot
      • TimeSeriesPredictionWithCI
      • TimeSeriesR2SquareBySegments
      • TokenDisparity
      • ToxicityScore
      • Embeddings
        • ClusterDistribution
        • CosineSimilarityComparison
        • CosineSimilarityDistribution
        • CosineSimilarityHeatmap
        • DescriptiveAnalytics
        • EmbeddingsVisualization2D
        • EuclideanDistanceComparison
        • EuclideanDistanceHeatmap
        • PCAComponentsPairwisePlots
        • StabilityAnalysisKeyword
        • StabilityAnalysisRandomNoise
        • StabilityAnalysisSynonyms
        • StabilityAnalysisTranslation
        • TSNEComponentsPairwisePlots
      • Ragas
        • AnswerCorrectness
        • AspectCritic
        • ContextEntityRecall
        • ContextPrecision
        • ContextPrecisionWithoutReference
        • ContextRecall
        • Faithfulness
        • NoiseSensitivity
        • ResponseRelevancy
        • SemanticSimilarity
      • Sklearn
        • AdjustedMutualInformation
        • AdjustedRandIndex
        • CalibrationCurve
        • ClassifierPerformance
        • ClassifierThresholdOptimization
        • ClusterCosineSimilarity
        • ClusterPerformanceMetrics
        • CompletenessScore
        • ConfusionMatrix
        • FeatureImportance
        • FowlkesMallowsScore
        • HomogeneityScore
        • HyperParametersTuning
        • KMeansClustersOptimization
        • MinimumAccuracy
        • MinimumF1Score
        • MinimumROCAUCScore
        • ModelParameters
        • ModelsPerformanceComparison
        • OverfitDiagnosis
        • PermutationFeatureImportance
        • PopulationStabilityIndex
        • PrecisionRecallCurve
        • RegressionErrors
        • RegressionErrorsComparison
        • RegressionPerformance
        • RegressionR2Square
        • RegressionR2SquareComparison
        • RobustnessDiagnosis
        • ROCCurve
        • ScoreProbabilityAlignment
        • SHAPGlobalImportance
        • SilhouettePlot
        • TrainingTestDegradation
        • VMeasure
        • WeakspotsDiagnosis
      • Statsmodels
        • AutoARIMA
        • CumulativePredictionProbabilities
        • DurbinWatsonTest
        • GINITable
        • KolmogorovSmirnov
        • Lilliefors
        • PredictionProbabilitiesHistogram
        • RegressionCoeffs
        • RegressionFeatureSignificance
        • RegressionModelForecastPlot
        • RegressionModelForecastPlotLevels
        • RegressionModelSensitivityPlot
        • RegressionModelSummary
        • RegressionPermutationFeatureImportance
        • ScorecardHistogram
    • Ongoing Monitoring
      • CalibrationCurveDrift
      • ClassDiscriminationDrift
      • ClassificationAccuracyDrift
      • ClassImbalanceDrift
      • ConfusionMatrixDrift
      • CumulativePredictionProbabilitiesDrift
      • FeatureDrift
      • PredictionAcrossEachFeature
      • PredictionCorrelation
      • PredictionProbabilitiesHistogramDrift
      • PredictionQuantilesAcrossFeatures
      • ROCCurveDrift
      • ScoreBandsDrift
      • ScorecardHistogramDrift
      • TargetPredictionDistributionPlot
    • Prompt Validation
      • Bias
      • Clarity
      • Conciseness
      • Delimitation
      • NegativeInstruction
      • Robustness
      • Specificity
  • Test sandbox beta

  • Notebooks
  • Code samples
    • Capital Markets
      • Quickstart for knockout option pricing model documentation
      • Quickstart for Heston option pricing model using QuantLib
    • Code Explainer
      • Quickstart for model code documentation
    • Credit Risk
      • Document an application scorecard model
      • Document an application scorecard model
      • Document a credit risk model
      • Document an application scorecard model
    • Custom Tests
      • Implement custom tests
      • Integrate external test providers
    • Model Validation
      • Validate an application scorecard model
    • Nlp and Llm
      • Sentiment analysis of financial data using a large language model (LLM)
      • Summarization of financial data using a large language model (LLM)
      • Sentiment analysis of financial data using Hugging Face NLP models
      • Summarization of financial data using Hugging Face NLP models
      • Automate news summarization using LLMs
      • Prompt validation for large language models (LLMs)
      • RAG Model Benchmarking Demo
      • RAG Model Documentation Demo
    • Ongoing Monitoring
      • Ongoing Monitoring for Application Scorecard
      • Quickstart for ongoing monitoring of models with ValidMind
    • Regression
      • Document a California Housing Price Prediction regression model
    • Time Series
      • Document a time series forecasting model
      • Document a time series forecasting model

  • Reference
  • ValidMind Library Python API
  • ​ValidMind Public REST API

On this page

  • ClassificationAccuracyDrift
    • Purpose
    • Test Mechanism
    • Signs of High Risk
    • Strengths
    • Limitations
  • Edit this page
  • Report an issue
  1. Test descriptions
  2. Ongoing Monitoring
  3. ClassificationAccuracyDrift

ClassificationAccuracyDrift

Compares classification accuracy metrics between reference and monitoring datasets.

Purpose

The Classification Accuracy Drift test is designed to evaluate changes in the model's predictive accuracy over time. By comparing key accuracy metrics between reference and monitoring datasets, this test helps identify whether the model maintains its performance levels in production. This is crucial for understanding if the model's predictions remain reliable and whether its overall effectiveness has degraded significantly.

Test Mechanism

This test proceeds by calculating comprehensive accuracy metrics for both reference and monitoring datasets. It computes overall accuracy, per-label precision, recall, and F1 scores, as well as macro-averaged metrics. The test quantifies drift as percentage changes in these metrics between datasets, providing both granular and aggregate views of accuracy changes. Special attention is paid to per-label performance to identify class-specific degradation.

Signs of High Risk

  • Large drifts in accuracy metrics exceeding the threshold
  • Inconsistent changes across different labels
  • Significant drops in macro-averaged metrics
  • Systematic degradation in specific class performance
  • Unexpected improvements suggesting data quality issues
  • Divergent trends between precision and recall

Strengths

  • Provides comprehensive accuracy assessment
  • Identifies class-specific performance changes
  • Enables early detection of model degradation
  • Includes both micro and macro perspectives
  • Supports multi-class classification evaluation
  • Maintains interpretable drift thresholds

Limitations

  • May be sensitive to class distribution changes
  • Does not account for prediction confidence
  • Cannot identify root causes of accuracy drift
  • Limited to accuracy-based metrics only
  • Requires sufficient samples per class
  • May not capture subtle performance changes
ClassDiscriminationDrift
ClassImbalanceDrift
  • ValidMind Logo
    ©
    Copyright 2025 ValidMind Inc.
    All Rights Reserved.
    Cookie preferences
    Legal
  • Get started
    • Model development
    • Model validation
    • Setup & admin
  • Guides
    • Access
    • Configuration
    • Model inventory
    • Model documentation
    • Model validation
    • Model workflows
    • Reporting
    • Monitoring
    • Attestation
  • Library
    • For developers
    • For validators
    • Code samples
    • API Reference
  • Training
    • Learning paths
    • Courses
    • Videos
  • Support
    • Troubleshooting
    • FAQ
    • Get help
  • Community
    • Slack
    • GitHub
    • Blog
  • Edit this page
  • Report an issue