Formula Used
MSE = (1 / n) × Σ(actual − predicted)²
Each prediction error is squared before averaging. Larger errors therefore receive greater weight.
RMSE is the square root of MSE. MAE averages absolute errors instead.
How to Use
- Paste curve values or actual and predicted observations.
- Select the curve mode and processing controls.
- Choose smoothing, normalization, outlier, and chart options.
- Configure early stopping and the monitored curve.
- Calculate, inspect diagnostics, and export the report.
Example Data Format
x,training_mse,validation_mse,testing_mse,model 1,0.92,1.02,1.06,Model A 2,0.70,0.79,0.84,Model A 3,0.53,0.61,0.66,Model A
Prediction format may use: x,actual,predicted,split,model.
Understanding the MSE Curve
A falling training curve usually indicates successful optimization. Validation behavior reveals whether improvements generalize.
A widening training-validation gap can indicate overfitting. A flat, high pair of curves often indicates underfitting.
MSE is sensitive to large errors and outliers. Always inspect the data scale and error distribution.
Frequently Asked Questions
What does a lower MSE mean?
A lower MSE means predictions are closer to actual values on average.
Can MSE equal zero?
Yes. Zero means every included prediction exactly matches its actual value.
Why does MSE emphasize outliers?
Squaring errors increases the contribution of unusually large prediction mistakes.
How is RMSE different?
RMSE returns the error to the original target unit.
When does the curve show overfitting?
Overfitting often appears when training MSE falls while validation MSE rises.
What is early stopping?
Early stopping ends training after monitored performance stops improving sufficiently.
Should MSE use a logarithmic axis?
A logarithmic axis helps when values span several orders of magnitude.
Can several models be compared?
Yes. Add a model column and provide rows for each model.
Should outliers always be removed?
No. Remove them only when justified by data quality or analysis goals.