generate_dataset.py

#!/usr/bin/env python3
"""
African Anti-Corruption Enforcement Dataset Generator

Generates synthetic records for 12 Sub-Saharan African countries across
three enforcement scenarios. Parameters are informed by:
  - Transparency International CPI 2025 (SSA avg=32, range 9-68)
  - EFCC Nigeria 2024 statistics (12,928 investigated, 5,083 prosecuted,
    4,111 convictions, ~$214M recovered)
  - South Africa SIU 2023/24 (1,919 investigations, 583 referrals,
    R8B saved, ~80% conviction rate in commercial cases)
  - Kenya EACC 2023/24 (5,171 reports, 534 under probe, 26.7% conviction
    rate, Ksh2.9B recovered)
  - AU Common African Position on Asset Recovery (asset recovery <10%)
  - G20 2025 Accountability Report on Whistleblower Protection
"""

import os
import argparse
import numpy as np
import pandas as pd
from scipy import stats

RNG_SEED = 42

# ── Country parameters ──────────────────────────────────────────────
# Based on TI CPI 2025, population, anti-corruption agency capacity,
# whistleblower protection maturity, and typical caseload volumes.
COUNTRIES = {
    "Nigeria": {
        "cpi_mean": 32, "cpi_std": 2.0,
        "pop_millions": 220,
        "agency_independence_mean": 42, "agency_independence_std": 8,
        "whistleblower_protection_mean": 33, "whistleblower_protection_std": 7,
        "cases_investigated_range": (10000, 18000),  # EFCC 12,928 in 2024
        "prosecution_rate_base": 0.35,  # EFCC: 5,083/12,928 ≈ 39%
        "conviction_rate_base": 0.72,   # EFCC high conviction capacity
        "asset_recovery_rate_base": 0.08,
        "assets_usd_millions_range": (80, 250),  # EFCC $214M in 2024
    },
    "South Africa": {
        "cpi_mean": 43, "cpi_std": 2.5,
        "pop_millions": 60,
        "agency_independence_mean": 55, "agency_independence_std": 10,
        "whistleblower_protection_mean": 48, "whistleblower_protection_std": 9,
        "cases_investigated_range": (1800, 2800),  # SIU 1,919 closed in 2023/24
        "prosecution_rate_base": 0.28,  # SIU 583 referrals from 1,919
        "conviction_rate_base": 0.75,   # NPA ~80% in commercial cases
        "asset_recovery_rate_base": 0.12,
        "assets_usd_millions_range": (100, 350),  # SIU R8B ≈ $430M
    },
    "Kenya": {
        "cpi_mean": 30, "cpi_std": 2.0,
        "pop_millions": 55,
        "agency_independence_mean": 38, "agency_independence_std": 7,
        "whistleblower_protection_mean": 30, "whistleblower_protection_std": 6,
        "cases_investigated_range": (4000, 6000),  # EACC 5,171 reports
        "prosecution_rate_base": 0.22,  # EACC forwards to ODPP
        "conviction_rate_base": 0.45,   # EACC 26.7% (2023/24), improving
        "asset_recovery_rate_base": 0.06,
        "assets_usd_millions_range": (15, 40),
    },
    "Ghana": {
        "cpi_mean": 43, "cpi_std": 2.5,
        "pop_millions": 34,
        "agency_independence_mean": 50, "agency_independence_std": 8,
        "whistleblower_protection_mean": 40, "whistleblower_protection_std": 7,
        "cases_investigated_range": (800, 1500),
        "prosecution_rate_base": 0.30,
        "conviction_rate_base": 0.55,
        "asset_recovery_rate_base": 0.07,
        "assets_usd_millions_range": (8, 30),
    },
    "Tanzania": {
        "cpi_mean": 38, "cpi_std": 2.0,
        "pop_millions": 65,
        "agency_independence_mean": 35, "agency_independence_std": 7,
        "whistleblower_protection_mean": 25, "whistleblower_protection_std": 6,
        "cases_investigated_range": (2000, 4000),
        "prosecution_rate_base": 0.25,
        "conviction_rate_base": 0.60,
        "asset_recovery_rate_base": 0.05,
        "assets_usd_millions_range": (10, 35),
    },
    "Uganda": {
        "cpi_mean": 26, "cpi_std": 2.0,
        "pop_millions": 48,
        "agency_independence_mean": 30, "agency_independence_std": 7,
        "whistleblower_protection_mean": 22, "whistleblower_protection_std": 5,
        "cases_investigated_range": (1500, 3000),
        "prosecution_rate_base": 0.18,
        "conviction_rate_base": 0.40,
        "asset_recovery_rate_base": 0.04,
        "assets_usd_millions_range": (5, 20),
    },
    "Ethiopia": {
        "cpi_mean": 30, "cpi_std": 2.5,
        "pop_millions": 120,
        "agency_independence_mean": 28, "agency_independence_std": 8,
        "whistleblower_protection_mean": 18, "whistleblower_protection_std": 5,
        "cases_investigated_range": (1000, 2500),
        "prosecution_rate_base": 0.15,
        "conviction_rate_base": 0.50,
        "asset_recovery_rate_base": 0.04,
        "assets_usd_millions_range": (5, 25),
    },
    "Senegal": {
        "cpi_mean": 46, "cpi_std": 2.0,
        "pop_millions": 18,
        "agency_independence_mean": 55, "agency_independence_std": 8,
        "whistleblower_protection_mean": 42, "whistleblower_protection_std": 7,
        "cases_investigated_range": (600, 1200),
        "prosecution_rate_base": 0.32,
        "conviction_rate_base": 0.55,
        "asset_recovery_rate_base": 0.08,
        "assets_usd_millions_range": (5, 20),
    },
    "Mozambique": {
        "cpi_mean": 21, "cpi_std": 2.0,
        "pop_millions": 33,
        "agency_independence_mean": 22, "agency_independence_std": 6,
        "whistleblower_protection_mean": 15, "whistleblower_protection_std": 4,
        "cases_investigated_range": (300, 800),
        "prosecution_rate_base": 0.12,
        "conviction_rate_base": 0.35,
        "asset_recovery_rate_base": 0.03,
        "assets_usd_millions_range": (2, 10),
    },
    "Zambia": {
        "cpi_mean": 33, "cpi_std": 2.0,
        "pop_millions": 20,
        "agency_independence_mean": 40, "agency_independence_std": 7,
        "whistleblower_protection_mean": 28, "whistleblower_protection_std": 6,
        "cases_investigated_range": (500, 1200),
        "prosecution_rate_base": 0.20,
        "conviction_rate_base": 0.45,
        "asset_recovery_rate_base": 0.05,
        "assets_usd_millions_range": (3, 15),
    },
    "Rwanda": {
        "cpi_mean": 58, "cpi_std": 2.5,
        "pop_millions": 14,
        "agency_independence_mean": 65, "agency_independence_std": 8,
        "whistleblower_protection_mean": 55, "whistleblower_protection_std": 8,
        "cases_investigated_range": (400, 900),
        "prosecution_rate_base": 0.45,
        "conviction_rate_base": 0.70,
        "asset_recovery_rate_base": 0.15,
        "assets_usd_millions_range": (3, 12),
    },
    "Botswana": {
        "cpi_mean": 58, "cpi_std": 2.5,
        "pop_millions": 2.6,
        "agency_independence_mean": 62, "agency_independence_std": 8,
        "whistleblower_protection_mean": 50, "whistleblower_protection_std": 8,
        "cases_investigated_range": (200, 500),
        "prosecution_rate_base": 0.40,
        "conviction_rate_base": 0.65,
        "asset_recovery_rate_base": 0.12,
        "assets_usd_millions_range": (2, 8),
    },
}

SCENARIOS = {
    "baseline": {
        "cpi_shift": 0,
        "agency_shift": 0,
        "whistleblower_shift": 0,
        "prosecution_multiplier": 1.0,
        "conviction_multiplier": 1.0,
        "asset_recovery_multiplier": 1.0,
        "investigated_multiplier": 1.0,
    },
    "strengthened_enforcement": {
        "cpi_shift": 8,          # +8 points on CPI
        "agency_shift": 18,       # +18 on independence score
        "whistleblower_shift": 20, # +20 on protection score
        "prosecution_multiplier": 1.45,
        "conviction_multiplier": 1.15,
        "asset_recovery_multiplier": 2.0,
        "investigated_multiplier": 1.25,
    },
    "weakened_accountability": {
        "cpi_shift": -10,
        "agency_shift": -15,
        "whistleblower_shift": -15,
        "prosecution_multiplier": 0.60,
        "conviction_multiplier": 0.80,
        "asset_recovery_multiplier": 0.40,
        "investigated_multiplier": 0.70,
    },
}


def clamp(x, lo, hi):
    return np.clip(x, lo, hi)


def generate_country_scenario(country_name, params, scenario_name, scenario_params, n_records, rng):
    """Generate n_records for one country under one scenario."""

    # ── Scores derived from scenario ──
    cpi = rng.normal(
        params["cpi_mean"] + scenario_params["cpi_shift"],
        params["cpi_std"], n_records
    )
    cpi = clamp(cpi, 0, 100)

    agency_independence = rng.normal(
        params["agency_independence_mean"] + scenario_params["agency_shift"],
        params["agency_independence_std"], n_records
    )
    agency_independence = clamp(agency_independence, 0, 100)

    whistleblower_protection = rng.normal(
        params["whistleblower_protection_mean"] + scenario_params["whistleblower_shift"],
        params["whistleblower_protection_std"], n_records
    )
    whistleblower_protection = clamp(whistleblower_protection, 0, 100)

    # ── Cases investigated (population-scaled) ──
    pop_scale = params["pop_millions"] / 60  # normalise to SA baseline
    inv_lo, inv_hi = params["cases_investigated_range"]
    inv_base = rng.uniform(inv_lo * pop_scale, inv_hi * pop_scale, n_records)
    cases_investigated = np.round(
        inv_base * scenario_params["investigated_multiplier"]
    ).astype(int)
    cases_investigated = np.maximum(cases_investigated, 10)

    # ── Prosecution rate (correlated with agency independence) ──
    prot_base = params["prosecution_rate_base"] * scenario_params["prosecution_multiplier"]
    # Agency independence boosts prosecution rate
    agency_effect = (agency_independence - 50) / 500  # ±0.1 shift
    prosecution_rate = prot_base + agency_effect + rng.normal(0, 0.04, n_records)
    prosecution_rate = clamp(prosecution_rate, 0.02, 0.85)

    cases_prosecuted = np.round(cases_investigated * prosecution_rate).astype(int)
    cases_prosecuted = np.minimum(cases_prosecuted, cases_investigated)
    cases_prosecuted = np.maximum(cases_prosecuted, 0)

    # ── Conviction rate ──
    conv_base = params["conviction_rate_base"] * scenario_params["conviction_multiplier"]
    conv_rate = conv_base + rng.normal(0, 0.06, n_records)
    conv_rate = clamp(conv_rate, 0.05, 0.95)

    convictions = np.round(cases_prosecuted * conv_rate).astype(int)
    convictions = np.minimum(convictions, cases_prosecuted)
    convictions = np.maximum(convictions, 0)

    # ── Asset confiscation ──
    asset_lo, asset_hi = params["assets_usd_millions_range"]
    assets_confiscated = rng.uniform(asset_lo, asset_hi, n_records) * scenario_params["asset_recovery_multiplier"]
    assets_confiscated = np.maximum(assets_confiscated, 0.1)

    asset_recovery_rate = params["asset_recovery_rate_base"] * scenario_params["asset_recovery_multiplier"]
    asset_recovery_rate = clamp(
        asset_recovery_rate + rng.normal(0, 0.02, n_records), 0.01, 0.50
    )

    # ── Whistleblower reports ──
    wb_base = cases_investigated * rng.uniform(0.3, 0.8, n_records)
    wb_effect = 1 + (whistleblower_protection - 30) / 100
    whistleblower_reports = np.round(wb_base * wb_effect).astype(int)
    whistleblower_reports = np.maximum(whistleblower_reports, 0)

    # ── Enforcement effectiveness composite ──
    enforcement_effectiveness = (
        0.25 * (prosecution_rate / 0.5) +
        0.25 * (conv_rate / 0.8) +
        0.20 * (asset_recovery_rate / 0.15) +
        0.15 * (agency_independence / 100) +
        0.15 * (whistleblower_protection / 100)
    ) * 100
    enforcement_effectiveness = clamp(enforcement_effectiveness, 0, 100)

    # ── Enforcement class ──
    enforcement_class = np.where(
        enforcement_effectiveness >= 60, "strong",
        np.where(enforcement_effectiveness >= 35, "moderate", "weak")
    )

    # ── Build DataFrame ──
    df = pd.DataFrame({
        "country": country_name,
        "scenario": scenario_name,
        "corruption_perception_index": np.round(cpi, 1),
        "cases_investigated": cases_investigated,
        "cases_prosecuted": cases_prosecuted,
        "prosecution_rate": np.round(prosecution_rate, 4),
        "convictions": convictions,
        "conviction_rate": np.round(conv_rate, 4),
        "assets_confiscated_usd_millions": np.round(assets_confiscated, 2),
        "asset_recovery_rate": np.round(asset_recovery_rate, 4),
        "whistleblower_reports": whistleblower_reports,
        "whistleblower_protection_score": np.round(whistleblower_protection, 1),
        "agency_independence_score": np.round(agency_independence, 1),
        "enforcement_effectiveness_score": np.round(enforcement_effectiveness, 1),
        "enforcement_class": enforcement_class,
    })
    return df


def generate_dataset(scenario, n_per_country=834, seed=RNG_SEED):
    """Generate full dataset for one scenario across all 12 countries."""
    rng = np.random.default_rng(seed)
    scenario_params = SCENARIOS[scenario]
    frames = []
    for country_name, params in COUNTRIES.items():
        df = generate_country_scenario(
            country_name, params, scenario, scenario_params, n_per_country, rng
        )
        frames.append(df)
    full = pd.concat(frames, ignore_index=True)
    # Shuffle
    full = full.sample(frac=1, random_state=seed).reset_index(drop=True)
    return full


def main():
    parser = argparse.ArgumentParser(description="Generate anti-corruption enforcement dataset")
    parser.add_argument("--scenario", type=str, default="all",
                        choices=list(SCENARIOS.keys()) + ["all"])
    parser.add_argument("--n-records", type=int, default=10000,
                        help="Total records per scenario (split across 12 countries)")
    parser.add_argument("--output-dir", type=str, default="data")
    args = parser.parse_args()

    os.makedirs(args.output_dir, exist_ok=True)
    scenarios = list(SCENARIOS.keys()) if args.scenario == "all" else [args.scenario]

    for scen in scenarios:
        n_per_country = args.n_records // len(COUNTRIES)
        remainder = args.n_records - n_per_country * len(COUNTRIES)
        df = generate_dataset(scen, n_per_country=n_per_country)
        # Add extra records to first country to hit exact count
        if remainder > 0:
            rng = np.random.default_rng(RNG_SEED + 99)
            extra = generate_country_scenario(
                list(COUNTRIES.keys())[0],
                COUNTRIES[list(COUNTRIES.keys())[0]],
                scen, SCENARIOS[scen], remainder, rng
            )
            df = pd.concat([df, extra], ignore_index=True)
            df = df.sample(frac=1, random_state=RNG_SEED).reset_index(drop=True)

        out_path = os.path.join(args.output_dir, f"{scen}.csv")
        df.to_csv(out_path, index=False)
        print(f"Generated {len(df)} records → {out_path}")


if __name__ == "__main__":
    main()