You or the network? A credit game
You are the credit officer of a bank. You see 10 real home-equity loan applications and decide: approve or refuse. After each decision you see what the network of Lecture 1 decided and whether the client repaid. In the full data 19.9% of the loans defaulted; in the game one application in three is a default.
The network is the multilayer perceptron (MLP 64-64) of the lecture, trained on 3,576 other loans; it never saw these applications. It refuses a loan when its probability of default is at least 50%.
What the game shows
- On all 1,192 test loans the network decides correctly in 92.1% of cases (AUC, the area under the ROC curve, 0.942).
- A signal people tend to overlook: when the debt-to-income ratio is missing from the file, 62.0% of the loans default; when it is reported, 8.6%. The network learned this from the data.
- 50% is not necessarily the right threshold: if a default costs five times as much as a lost good client, the cost-minimising threshold is 1/6 (slide 40 of the lecture).
Data: HMEQ home-equity loans (Baesens, Roesch and Scheule, 2016; 5,960 loans). The 36 applications of the game are drawn at random from the test loans (12 defaults, 24 repaid; seed 20261108); every game draws 10 of them. Code: lectures/L1/make_game_L1.py.

