Section 1: Model structure

```

[ Input vector x in ℝ⁸ ]
(F3_BM, F3_BI, F3_Math, F3_Science, F3_Sejarah, F3_Geo, F3_RBT, F3_PSV)
│
▼
[ StandardScaler ]
z = (x - mean) / std   (mean and std from training stream only)
│
▼
[ Ridge Regression ]
y_hat = intercept + Σ(coef_j * z_j)
│
▼
[ Output ]
Predicted merit (float, clipped to 0-100)


```
  
Section 4: Evaluation
  

    Model selection: 5-fold KFold (shuffle, random_state=42). Metric: mean absolute error (MAE) on held-out folds. Full metrics per stream and per candidate are stored in model_meta.pkl. Separate test set not used because sample size per stream is small.
  

  
Section 5: Limitations
  

    
Trained only on students who chose each stream (selection bias, extrapolation)
    
Small per-stream sample size
    
Predicts an estimate, not a guarantee
    
Model version: scikit-learn 1.6.1 (backend environment must match)
  

Section 2: Coefficient table

(standardised features; coefficient = change in merit per 1 standard deviation change in that subject)

Caption: Negative coefficients on correlated subjects are typical multicollinearity effects. Do not interpret individually as causal.

Hyperparameters: alpha = 5.0 (all three streams)

SCIENCE

SEMI SCIENCE

ART