Analysis results
Results update without reloading the page.
Interpretation
Series ranking
| Rank | Series | Minimum | Final | Average | Stability |
|---|---|---|---|---|---|
| Calculate results to view rankings. | |||||
Detailed results
| No calculated rows yet. |
Formula used
An error rate measures the proportion of predictions that are wrong. Lower values usually indicate stronger predictive performance. Context still matters when classes are imbalanced.
The monitored curve determines the selected best step. Validation error is commonly used for model selection. Testing data should remain untouched during tuning.
The generalization gap compares validation and training behavior. A widening positive gap can indicate overfitting. Stable curves often suggest reliable convergence.
How to use this calculator
- Select the input value format and monitored error series.
- Enter rows manually or import CSV data with headers.
- Choose smoothing, normalization, chart, and early-stopping options.
- Select Calculate and plot to generate the analysis.
- Review the best step, rankings, insights, and detailed table.
- Export the chart, CSV data, PDF report, or copied summary.
Example data
| Epoch | Training error | Validation error | Testing error |
|---|---|---|---|
| 1 | 0.285 | 0.310 | 0.325 |
| 2 | 0.245 | 0.274 | 0.291 |
| 3 | 0.208 | 0.239 | 0.258 |
| 4 | 0.176 | 0.212 | 0.229 |
| 5 | 0.149 | 0.194 | 0.209 |
| 6 | 0.126 | 0.181 | 0.198 |
Understanding error rate curves
Training versus validation error
Training error describes performance on examples used during optimization. Validation error estimates performance on unseen tuning data. Their difference reveals useful generalization behavior.
Overfitting and underfitting
Overfitting often appears when training error keeps falling. Validation error may then flatten or increase. Underfitting keeps both curves undesirably high.
Early stopping
Early stopping watches a selected validation metric. Training stops after improvement remains insufficient for several steps. Patience reduces reactions to temporary noise.
Curve smoothing
Smoothing makes noisy trends easier to inspect. It should not replace raw metric reporting. Compare smoothed and unsmoothed behavior carefully.
Frequently asked questions
What is a good error rate?
A good value depends on the task, baseline, class balance, and consequences of mistakes. Compare models using consistent data and evaluation procedures.
Should I monitor training or validation error?
Validation error is usually better for model selection and early stopping. Training error mainly shows optimization progress.
Why can validation error increase?
The model may be overfitting, the validation sample may be noisy, or the optimization process may be unstable.
What is balanced error rate?
Balanced error rate averages false-positive and false-negative rates. It can be more informative for imbalanced classification problems.
Can percentages and decimals be mixed?
Choose the matching input format before calculating. Percentage mode treats 12 as 12 percent, while decimal mode treats 0.12 as twelve percent.
What does the smoothing window control?
It controls how many nearby values influence moving-average or rolling-median results. Larger windows create smoother but less responsive curves.
How is the best step selected?
The calculator finds the minimum processed value in the monitored series. The corresponding X-axis value becomes the best step.
Does early stopping guarantee the best model?
No. It is a practical rule that limits unnecessary training. Final model choice should also consider stability and external testing.
Why use a logarithmic Y-axis?
A logarithmic axis can reveal proportional changes across wide ranges. It cannot display zero or negative values.