Lab notebook · 14 / 27Risk modelling

Credit scoring with self-organizing maps

What structure does a self-organizing map reveal, and how do supervised baselines perform on held-out credit records?

Self-directed study

Scope of this result

One split of a small historical dataset. The result does not establish fairness or suitability for lending decisions.

The evidence

Fig. 01

Two heatmaps of a 10 by 10 self-organizing map showing training bad-credit share and number of training rows assigned to each node. — The training-set SOM shows node-level bad-credit share beside node occupancy; empty nodes are left blank.

01 · Method and interpretation

A 10×10 self-organizing map explores profiles in UCI’s corrected South German Credit data, followed by supervised baselines on one stratified held-out split. The same training-fitted scaling is applied to test rows, and the personal-status/sex and foreign-worker fields are excluded. Balanced logistic regression reaches 75.2% balanced accuracy and 78.3% bad-risk recall, but this small historical dataset and one-split result are not suitable for lending decisions or evidence of fairness.

02 · Supporting figures

Selected outputs from the complete notebook.

Fig. 02

Confusion matrix with 47 bad-risk and 101 good-risk test rows correctly classified, alongside 13 and 39 errors. — Held-out confusion matrix for balanced logistic regression: 47 of 60 bad-risk rows and 101 of 140 good-risk rows classified correctly.