Skip to main content

Model Performance Metrics

RiverGen Model Studio provides a standardized set of metrics to evaluate the performance of your machine learning models. These metrics are automatically calculated during the Validation Phase of the training pipeline and are synchronized via the Metrics Synchronization Worker.


1. Classification Metrics

Used for tasks where the target is a category (e.g., fraud vs. legitimate).

MetricDescriptionIdeal ValueUse Case
AccuracyProportion of total predictions that were correct.1.0Balanced datasets.
PrecisionRatio of correct positive observations to total predicted positives.1.0When false positives are costly (e.g., spam detection).
RecallRatio of correct positive observations to actual positives.1.0When false negatives are costly (e.g., disease detection).
F1-ScoreWeighted average of Precision and Recall.1.0Imbalanced datasets.
Log LossMeasure of uncertainty of your predictions.0.0Probability-based classification.

2. Regression Metrics

Used for tasks where the target is a continuous number (e.g., resource utilization).

MetricDescriptionIdeal ValueUse Case
MAEMean Absolute Error. Average magnitude of the errors.0.0Easy to interpret in units of the target.
MSEMean Squared Error. Penalizes large errors more heavily.0.0Sensitive to outliers.
RMSERoot Mean Squared Error. Standard deviation of residuals.0.0Standard metric for error magnitude.
R² (R-Squared)Proportion of variance explained by the model.1.0Measuring model fit.
Adjusted R²R-Squared adjusted for the number of predictors.1.0Comparing models with different numbers of features.

3. Time-Series Metrics

Used for forecasting and sequential data modeling.

MetricDescriptionIdeal ValueUse Case
MAPEMean Absolute Percentage Error.0.0%Relative error comparison.
SMAPESymmetric MAPE.0.0%Handling zero values in data.
MASEMean Absolute Scaled Error.< 1.0Comparing different time series.

4. Interpretation Guide

Confusion Matrix

For classification tasks, Model Studio generates a Confusion Matrix to visualize the performance of an algorithm.

Residual Analysis

For regression tasks, residuals are plotted to identify patterns that the model failed to capture. A good model should have residuals randomly distributed around zero.


Advanced: Custom Metrics

If you are using Custom Model Training, you can emit your own metrics by writing a JSON file to the RG_OUTPUT_PATH. These will be automatically parsed and mapped to the Model Metrics table.

{
"custom_metric_name": "Stability Score",
"value": 0.942,
"timestamp": "2026-02-27T10:00:00Z"
}
Automatic Sync

Metrics are synchronized every 30 seconds during training and stored permanently in the model_metrics table for longitudinal tracking.