generate_synthetic_data.py

"""
Generates SYNTHETIC credit-application data with a KNOWN ground-truth
relationship between features and outcome. This is deliberate: because we
know the true generating weights, we can later verify that the model
recovers a sensible approximation of them, and that the explanation layer
(SHAP-style + LIME-style) correctly attributes the decision to the features
that actually drove it -- a property you can't verify against messy real
data. No real applicant, credit bureau, or financial data is used.
"""
import csv
import os
import random

random.seed(7)

N = 600

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

# True (synthetic) generating weights -- unknown to the model at training time,
# used only to label the data and later sanity-check what the model/explainer recovers.
TRUE_WEIGHTS = {
    "annual_income_k": 0.028,
    "debt_to_income_pct": -0.055,
    "credit_history_years": 0.09,
    "num_late_payments_2y": -0.65,
    "employment_years": 0.05,
    "loan_amount_k": -0.012,
    "num_open_accounts": -0.04,
}
TRUE_INTERCEPT = -1.0


def sigmoid(x):
    import math
    return 1 / (1 + math.exp(-x))


def make_row():
    income = max(18, random.gauss(65, 28))
    dti = max(2, min(70, random.gauss(28, 12)))
    history = max(0, random.gauss(9, 6))
    late = max(0, int(random.gauss(1.1, 1.4)))
    employment = max(0, random.gauss(6, 5))
    loan = max(3, random.gauss(18, 12))
    open_accts = max(0, int(random.gauss(5, 2.5)))

    features = {
        "annual_income_k": round(income, 1),
        "debt_to_income_pct": round(dti, 1),
        "credit_history_years": round(history, 1),
        "num_late_payments_2y": late,
        "employment_years": round(employment, 1),
        "loan_amount_k": round(loan, 1),
        "num_open_accounts": open_accts,
    }

    logit = TRUE_INTERCEPT + sum(TRUE_WEIGHTS[k] * v for k, v in features.items())
    logit += random.gauss(0, 0.4)  # irreducible noise
    p_approve = sigmoid(logit)
    approved = 1 if random.random() < p_approve else 0

    features["approved"] = approved
    return features


def main():
    rows = [make_row() for _ in range(N)]
    out_path = os.path.join(os.path.dirname(__file__), "data", "credit_applications.csv")
    with open(out_path, "w", newline="") as f:
        writer = csv.DictWriter(f, fieldnames=FEATURES + ["approved"])
        writer.writeheader()
        writer.writerows(rows)
    approve_rate = sum(r["approved"] for r in rows) / len(rows)
    print(f"Wrote {len(rows)} synthetic applications to {out_path} (approve rate: {approve_rate:.1%})")


if __name__ == "__main__":
    main()