model.py

"""
The production model being audited: logistic regression over 7 synthetic
credit-application features. Logistic regression is a deliberate choice for
a compliance dashboard, not just a simplification -- it's the closed form
that makes exact, closed-form attribution possible (see explain/shap_exact.py)
rather than requiring a sampling-based approximation. In production this slot
is an arbitrary model (XGBoost, a neural net); the explanation layer in this
repo is built to swap over to KernelSHAP/LIME without assuming linearity --
see the production upgrade path in the README.
"""
from __future__ import annotations
import os
from dataclasses import dataclass

import numpy as np
import pandas as pd
from sklearn.linear_model import LogisticRegression

FEATURES = [
    "annual_income_k", "debt_to_income_pct", "credit_history_years",
    "num_late_payments_2y", "employment_years", "loan_amount_k", "num_open_accounts",
]

FEATURE_LABELS = {
    "annual_income_k": "Annual income ($k)",
    "debt_to_income_pct": "Debt-to-income (%)",
    "credit_history_years": "Credit history (years)",
    "num_late_payments_2y": "Late payments (last 2y)",
    "employment_years": "Employment length (years)",
    "loan_amount_k": "Loan amount requested ($k)",
    "num_open_accounts": "Open credit accounts",
}


@dataclass
class TrainedModel:
    model: LogisticRegression
    feature_means: pd.Series
    train_accuracy: float
    coefficients: dict


def train(csv_path: str) -> TrainedModel:
    df = pd.read_csv(csv_path)
    X = df[FEATURES]
    y = df["approved"]

    model = LogisticRegression(max_iter=1000)
    model.fit(X, y)

    acc = float(model.score(X, y))
    coefs = dict(zip(FEATURES, model.coef_[0]))

    return TrainedModel(
        model=model,
        feature_means=X.mean(),
        train_accuracy=acc,
        coefficients=coefs,
    )


def predict_proba(trained: TrainedModel, applicant: dict) -> float:
    x = pd.DataFrame([applicant])[FEATURES]
    return float(trained.model.predict_proba(x)[0][1])