Prediction inputs
Formula used
ŷ = F₀(x) + Σ[ηₘ × wₘ × hₘ(x)]
F(x) = F₀(x) + Σ[ηₘ × wₘ × hₘ(x)]
P(y = 1) = 1 ÷ (1 + e−F(x))
Fk(x) = F0,k(x) + Σ[ηₘ × wₘ × hm,k(x)]
Pk = eFk ÷ Σ eFj
How to use the calculator
- Select regression, binary classification, or multiclass classification.
- Enter the model’s base estimate, raw score, or base probabilities.
- Supply the global learning rate and sequential tree outputs.
- Enable per-tree rates or weights when your model uses them.
- Add an actual target or class to calculate error metrics.
- Submit the form, inspect staged results, and export the report.
Example data
| Mode | Base value | Learning rate | Tree outputs | Expected output type |
|---|---|---|---|---|
| Regression | 10 | 0.10 | 2.0, −1.0, 0.5 | 10.15 numeric prediction |
| Binary | 0 raw score | 0.10 | 0.8, 0.4, −0.2 | Probability and class label |
| Multiclass | 0, 0, 0 | 0.10 | One three-value vector per tree | Softmax probabilities |
Frequently asked questions
What is a gradient boosting prediction?
It combines a base estimate with many small tree corrections. Each tree updates the running score sequentially. The final score becomes the model prediction.
What does the learning rate do?
The learning rate shrinks every tree contribution. Smaller values usually produce more gradual updates. They often require additional boosting trees.
Can tree outputs be negative?
Yes, negative outputs reduce the current prediction score. Positive outputs increase the score instead. Both directions are normal in boosting.
What is the binary raw score?
It is the additive score before logistic transformation. It is often called a logit. The sigmoid converts it into probability.
How is a class selected?
Binary mode compares probability with your threshold. Multiclass mode selects the largest softmax probability. Labels make the decision easier to read.
Why use per-tree learning rates?
Some custom ensembles apply different shrinkage values. Standard libraries usually use one global rate. This option supports advanced manual calculations.
What does a tree weight represent?
A weight scales one tree before accumulation. A value of one changes nothing. Other values increase or reduce its influence.
Which error metrics are included?
Regression provides residual, absolute error, and squared error. Classification can calculate probability residuals and log loss. Actual values remain optional.
Does this calculator train a model?
No, it evaluates supplied boosting outputs only. Training requires data and tree-fitting algorithms. Use this tool for checking model calculations.