Boosting Loss Curve Calculator

Boosting Loss Curve Results Example data loaded

Review the example result, change options, then select Calculate Curve.

Model diagnosis

Analysis will appear here.

Recommended action

Recommendations will appear here.

Model comparison

Rank	Series	Minimum loss	Best round	Final loss	Reduction	Stability

1. Loss Data

CSV or spreadsheet data

Use one iteration column and any number of loss columns. Commas, tabs, and semicolons are accepted.

Data actions

Iteration column

Primary training series

Primary validation series

Confidence lower-bound series

Confidence upper-bound series

Manual stopping round

2. Model and Loss Configuration

Boosting model

Loss function

Horizontal axis meaning

Learning rate

Maximum tree depth

Regularisation strength

Subsample ratio

Feature sampling ratio

Target loss reference

3. Smoothing and Detection

Smoothing method

Smoothing window

EMA smoothing factor

Early-stopping patience

Minimum improvement

Overfitting confirmation rounds

Underfitting final-loss threshold

Displayed decimal places

Process from iteration

Process through iteration

Show original and smoothed curves

Detect early stopping

Detect overfitting

Normalize every loss series

Interpolate missing values

Remove duplicate iterations

Winsorize extreme outliers

4. Chart Configuration

Chart title

X-axis title

Y-axis title

Curve display

Line shape

Line width

Marker size

Legend position

Y-axis scale

Chart height

X-axis minimum

X-axis maximum

Y-axis minimum

Y-axis maximum

Show gridlines

Reverse loss axis

Mark best round

Show target-loss line

Highlight overfitting region

Connect data gaps

Show supplied confidence band

Custom annotation round

Custom annotation text

Custom curve names

Enter one original-column=name mapping per line.

Formula Used

Loss reduction = Initial loss − Final loss Percentage reduction = ((Initial loss − Final loss) / Initial loss) × 100 Generalisation gap = Validation loss − Training loss Average reduction per round = (Initial loss − Final loss) / Number of transitions Area under loss curve = Σ ((Lᵢ + Lᵢ₊₁) / 2) × (xᵢ₊₁ − xᵢ) Stability score = 100 / (1 + 25 × standard deviation of consecutive loss changes)

Early stopping tracks the lowest validation loss. Training stops after patience rounds without sufficient improvement. Overfitting begins when validation loss rises while training loss continues falling.

How to Use

Paste loss data or import a CSV file.
Select iteration, training, and validation columns.
Choose the boosting model and loss function.
Configure smoothing, early stopping, and detection thresholds.
Adjust chart labels, scale, markers, and annotations.
Select Calculate Curve to generate the analysis.
Export data, images, PDF reports, or copied results.

Example Data

Round	Training loss	Validation loss	Test loss
1	0.720	0.760	0.770
5	0.480	0.520	0.535
10	0.310	0.360	0.372
15	0.220	0.295	0.308
20	0.165	0.310	0.322

Frequently Asked Questions

What does a boosting loss curve show?

It shows how model error changes across boosting rounds. Training and validation curves reveal learning progress. Their separation can expose overfitting.

Which curve determines the best round?

Validation loss usually determines the best round. Choose its lowest reliable point. Testing data should remain untouched.

Why does training loss keep decreasing?

Each new learner corrects previous model errors. Training performance therefore often improves continuously. Validation performance may eventually worsen.

What is early stopping patience?

Patience counts tolerated non-improving rounds. Larger values allow more recovery time. Smaller values stop training sooner.

Should loss curves be smoothed?

Smoothing helps reveal broad trends. Excessive smoothing can hide instability. Compare original and smoothed curves together.

When is logarithmic scaling useful?

Logarithmic scaling helps with wide loss ranges. It exposes proportional changes more clearly. All displayed losses must remain positive.

How is overfitting detected?

Validation loss rises while training loss falls. The calculator checks consecutive confirming rounds. It then highlights the suspected region.

Can several models be compared?

Yes, add multiple numeric loss columns. Every series appears on the chart. Validation-like columns receive comparison rankings.

What causes unstable loss curves?

Learning rates may be too high. Samples or features may vary excessively. Noisy validation sets can also contribute.