Credit Scoring Model Trainer: Watch Logistic Regression Learn (2D)

Training Controls

Learning rate (η) 0.60

L2 regularization (λ) 0.05

Training speed 20 steps/s

Live Readout

Iteration

Loss (cross-entropy)

—

Training accuracy

—

Weights (w₁, w₂, b)

—

How it works

140 synthetic applicants are placed in a 2-feature space (normalized income, normalized debt-to-income ratio). Each carries a hidden "will default" label generated from a true — but noisy — risk rule the model never sees directly. A logistic regression classifier is trained live with full-batch gradient descent on the binary cross-entropy loss, plus an L2 penalty:

z  = w1·x1 + w2·x2 + b
p  = σ(z) = 1 / (1 + e^-z)
L  = −mean[ y·ln(p) + (1−y)·ln(1−p) ] + (λ/2)·(w1²+w2²)

∂L/∂w1 = mean[(p−y)·x1] + λ·w1
∂L/∂w2 = mean[(p−y)·x2] + λ·w2
∂L/∂b  = mean[(p−y)]

w ← w − η·∂L/∂w   (each training step)

The decision boundary — the line where p = 0.5, i.e. w1·x1 + w2·x2 + b = 0 — is redrawn after every step so you can watch it rotate and slide into place as the weights converge. A high learning rate converges fast but can oscillate or diverge; a high λ shrinks the weights and flattens the boundary, trading fit for a simpler, more conservative model.

Left pane — applicants in feature space, colored by their true label, shaded regions show the model's current approve/deny call.
Right pane — the loss curve over training iterations; it should fall monotonically toward a floor set by how separable the two classes are.

Repaid (true label) Defaulted (true label) Decision boundary / loss