generate_dataset.py

#!/usr/bin/env python3
"""Generate synthetic African crop insurance payout dataset.

Parameters derived from IBLI (Kenya/Ethiopia), WFP R4, ACRE Africa
programs and published literature on index-based insurance in SSA.
"""

import csv
import json
import os
import random
from datetime import datetime

import numpy as np

# --- Country parameters (calibrated from IBLI/R4/ACRE field data) ---
COUNTRIES = {
    "Kenya": {
        "regions": 8,
        "base_enrollment_rate": 0.12,
        "premium_rate": 0.055,
        "avg_premium_usd": 12,
        "drought_prob": 0.25,
        "ndvi_baseline": 0.48,
        "rainfall_baseline_mm": 650,
    },
    "Ethiopia": {
        "regions": 10,
        "base_enrollment_rate": 0.08,
        "premium_rate": 0.045,
        "avg_premium_usd": 8,
        "drought_prob": 0.30,
        "ndvi_baseline": 0.42,
        "rainfall_baseline_mm": 580,
    },
    "Tanzania": {
        "regions": 7,
        "base_enrollment_rate": 0.06,
        "premium_rate": 0.05,
        "avg_premium_usd": 10,
        "drought_prob": 0.20,
        "ndvi_baseline": 0.50,
        "rainfall_baseline_mm": 720,
    },
    "Malawi": {
        "regions": 5,
        "base_enrollment_rate": 0.09,
        "premium_rate": 0.06,
        "avg_premium_usd": 7,
        "drought_prob": 0.22,
        "ndvi_baseline": 0.46,
        "rainfall_baseline_mm": 900,
    },
    "Zambia": {
        "regions": 6,
        "base_enrollment_rate": 0.07,
        "premium_rate": 0.05,
        "avg_premium_usd": 9,
        "drought_prob": 0.18,
        "ndvi_baseline": 0.52,
        "rainfall_baseline_mm": 850,
    },
    "Zimbabwe": {
        "regions": 5,
        "base_enrollment_rate": 0.05,
        "premium_rate": 0.06,
        "avg_premium_usd": 8,
        "drought_prob": 0.28,
        "ndvi_baseline": 0.44,
        "rainfall_baseline_mm": 600,
    },
    "Mozambique": {
        "regions": 5,
        "base_enrollment_rate": 0.04,
        "premium_rate": 0.055,
        "avg_premium_usd": 6,
        "drought_prob": 0.22,
        "ndvi_baseline": 0.53,
        "rainfall_baseline_mm": 950,
    },
    "Senegal": {
        "regions": 4,
        "base_enrollment_rate": 0.06,
        "premium_rate": 0.05,
        "avg_premium_usd": 11,
        "drought_prob": 0.35,
        "ndvi_baseline": 0.35,
        "rainfall_baseline_mm": 450,
    },
    "Ghana": {
        "regions": 5,
        "base_enrollment_rate": 0.05,
        "premium_rate": 0.045,
        "avg_premium_usd": 10,
        "drought_prob": 0.15,
        "ndvi_baseline": 0.55,
        "rainfall_baseline_mm": 1100,
    },
    "Nigeria": {
        "regions": 8,
        "base_enrollment_rate": 0.03,
        "premium_rate": 0.05,
        "avg_premium_usd": 9,
        "drought_prob": 0.20,
        "ndvi_baseline": 0.50,
        "rainfall_baseline_mm": 1050,
    },
    "Rwanda": {
        "regions": 4,
        "base_enrollment_rate": 0.10,
        "premium_rate": 0.04,
        "avg_premium_usd": 7,
        "drought_prob": 0.12,
        "ndvi_baseline": 0.56,
        "rainfall_baseline_mm": 1200,
    },
    "Uganda": {
        "regions": 5,
        "base_enrollment_rate": 0.07,
        "premium_rate": 0.05,
        "avg_premium_usd": 8,
        "drought_prob": 0.18,
        "ndvi_baseline": 0.52,
        "rainfall_baseline_mm": 1150,
    },
}

CROP_TYPES = ["maize", "wheat", "sorghum", "rice", "cassava"]
CROP_WEIGHTS = [0.35, 0.15, 0.20, 0.15, 0.15]

INSURANCE_TYPES = ["index_based", "area_yield", "weather_derivative"]
INSURANCE_WEIGHTS = [0.50, 0.30, 0.20]

SEASONS = ["long_rains", "short_rains"]

SCENARIOS = {
    "baseline": {
        "enrollment_mult": 1.0,
        "payout_mult": 1.0,
        "drought_shift": 0.0,
        "description": "Current-state parameters based on existing IBLI/R4/ACRE programs",
    },
    "expanded_index_insurance": {
        "enrollment_mult": 2.5,
        "payout_mult": 1.1,
        "drought_shift": 0.0,
        "description": "Scaled enrollment via subsidy and mobile distribution (ACRE model)",
    },
    "climate_shock": {
        "enrollment_mult": 1.1,
        "payout_mult": 1.0,
        "drought_shift": 0.15,
        "description": "Increased drought frequency reflecting IPCC AR6 projections for SSA",
    },
}

# IBLI basis risk: Jensen et al. 2016 found ~69% residual risk exposure
# and ~63% reduction in covariate risk from high-loss events
BASIS_RISK_PARAMS = {
    "index_based": {"mean": 0.31, "std": 0.12},
    "area_yield": {"mean": 0.22, "std": 0.10},
    "weather_derivative": {"mean": 0.18, "std": 0.09},
}

YEARS = list(range(2015, 2025))

# Crop-specific yield parameters (tonnes/ha)
CROP_YIELD = {
    "maize": {"mean": 1.8, "std": 0.6},
    "wheat": {"mean": 1.5, "std": 0.5},
    "sorghum": {"mean": 0.9, "std": 0.4},
    "rice": {"mean": 2.2, "std": 0.7},
    "cassava": {"mean": 8.0, "std": 3.0},
}


def generate_record(record_id, country, params, scenario_name, scenario, year, season):
    """Generate a single insurance record."""
    rng = np.random.default_rng(seed=(record_id * 31 + hash(country) + year) % (2**31))

    # Drought probability with scenario shift
    drought_prob = min(params["drought_prob"] + scenario["drought_shift"], 0.75)
    is_drought = rng.random() < drought_prob

    # NDVI anomaly (z-score). Negative during drought
    # IBLI uses cumulative NDVI deviation; payout triggers at 15th percentile
    if is_drought:
        ndvi_anomaly = round(rng.normal(-1.2, 0.5), 3)
        ndvi_anomaly = max(ndvi_anomaly, -3.0)
    else:
        ndvi_anomaly = round(rng.normal(0.1, 0.4), 3)
        ndvi_anomaly = max(min(ndvi_anomaly, 2.0), -0.5)

    # Rainfall deficit percentage
    if is_drought:
        rainfall_deficit = round(rng.uniform(25, 65), 1)
    else:
        rainfall_deficit = round(max(0, rng.normal(5, 15)), 1)

    # Yield loss percentage - correlated with NDVI anomaly and rainfall deficit
    base_yield_loss = max(0, -ndvi_anomaly * 15 + rainfall_deficit * 0.3 + rng.normal(0, 5))
    yield_loss_pct = round(min(base_yield_loss, 90), 1)

    # Insurance type
    insurance_type = rng.choice(INSURANCE_TYPES, p=INSURANCE_WEIGHTS)

    # Crop type
    crop_type = rng.choice(CROP_TYPES, p=CROP_WEIGHTS)

    # Enrolled hectares (smallholder: 0.5-5 ha typical, ACRE data)
    enrolled_hectares = round(rng.lognormal(mean=1.2, sigma=0.7), 2)
    enrolled_hectares = max(0.2, min(enrolled_hectares, 20.0))

    # Premium paid (ACRE: 5-25% of harvest value; R4: $7-12 avg)
    base_premium = params["avg_premium_usd"]
    premium_paid = round(base_premium * enrolled_hectares * rng.uniform(0.8, 1.3), 2)

    # Payout determination
    # Index-based: triggers on NDVI anomaly below ~15th percentile (z < -1.0)
    # Area-yield: triggers when yield_loss_pct > 20%
    # Weather derivative: triggers on rainfall_deficit > 30%
    if insurance_type == "index_based":
        payout_triggered = ndvi_anomaly < -1.0
    elif insurance_type == "area_yield":
        payout_triggered = yield_loss_pct > 20.0
    else:
        payout_triggered = rainfall_deficit > 30.0

    # Apply scenario multiplier to slightly increase payout probability
    scenario_boost = 0.1 * (scenario["payout_mult"] - 1.0)
    if not payout_triggered and rng.random() < scenario_boost:
        payout_triggered = True

    # Payout amount
    # IBLI: payout proportional to severity above trigger
    # ACRE: payout ~ 60-90% of insured value when triggered
    if payout_triggered:
        severity = max(min(yield_loss_pct / 50.0, 1.0), 0.15)
        payout_ratio = round(max(rng.uniform(0.4, 0.9) * severity, 0.05), 3)
        insured_value = enrolled_hectares * CROP_YIELD[crop_type]["mean"] * 350  # USD/tonne
        payout_amount = round(max(insured_value * payout_ratio, 1.0), 2)
    else:
        payout_ratio = 0.0
        payout_amount = 0.0

    # Farmer count per record (represents a micro-insurance cohort)
    farmer_count = int(rng.lognormal(mean=4.5, sigma=1.0))
    farmer_count = max(5, min(farmer_count, 500))

    # Enrollment rate (scenario-adjusted)
    base_rate = params["base_enrollment_rate"]
    enrollment_rate = round(
        min(base_rate * scenario["enrollment_mult"] * rng.uniform(0.7, 1.3), 0.85), 4
    )

    # Lapse rate (IBLI: ~15-25% non-renewal; higher when no payout)
    if payout_triggered:
        lapse_rate = round(rng.uniform(0.05, 0.15), 4)
    else:
        lapse_rate = round(rng.uniform(0.15, 0.30), 4)

    # Basis risk score (0=perfect hedging, 1=no hedging)
    # Jensen et al. 2016: IBLI leaves ~69% residual risk
    br = BASIS_RISK_PARAMS[insurance_type]
    basis_risk_score = round(np.clip(rng.normal(br["mean"], br["std"]), 0.0, 1.0), 3)

    return {
        "record_id": record_id,
        "country": country,
        "year": year,
        "season": season,
        "crop_type": crop_type,
        "insurance_type": insurance_type,
        "scenario": scenario_name,
        "enrolled_hectares": enrolled_hectares,
        "premium_paid_usd": premium_paid,
        "ndvi_anomaly": ndvi_anomaly,
        "rainfall_deficit_pct": rainfall_deficit,
        "yield_loss_pct": yield_loss_pct,
        "payout_triggered": payout_triggered,
        "payout_amount_usd": payout_amount,
        "payout_ratio": payout_ratio,
        "farmer_count": farmer_count,
        "enrollment_rate": enrollment_rate,
        "lapse_rate": lapse_rate,
        "basis_risk_score": basis_risk_score,
    }


def generate_dataset():
    """Generate full dataset: 12 countries × 3 scenarios × 10000 records."""
    all_records = []
    record_id = 1

    for scenario_name, scenario in SCENARIOS.items():
        records_per_country = 10000 // len(COUNTRIES)
        remainder = 10000 % len(COUNTRIES)

        for i, (country, params) in enumerate(COUNTRIES.items()):
            n = records_per_country + (1 if i < remainder else 0)

            for _ in range(n):
                year = random.choice(YEARS)
                season = random.choice(SEASONS)
                rec = generate_record(
                    record_id, country, params, scenario_name, scenario, year, season
                )
                all_records.append(rec)
                record_id += 1

    return all_records


def write_csv(records, output_path):
    """Write records to CSV."""
    fieldnames = list(records[0].keys())
    with open(output_path, "w", newline="") as f:
        writer = csv.DictWriter(f, fieldnames=fieldnames)
        writer.writeheader()
        writer.writerows(records)


def write_metadata(output_dir):
    """Write dataset metadata JSON."""
    metadata = {
        "name": "african-crop-insurance-payouts",
        "description": (
            "Synthetic dataset of agricultural index-insurance payouts across "
            "12 Sub-Saharan African countries. Parameters calibrated from IBLI "
            "(Kenya/Ethiopia), WFP R4, and ACRE Africa programs."
        ),
        "version": "1.0.0",
        "created": datetime.utcnow().isoformat() + "Z",
        "license": "CC-BY-4.0",
        "scenarios": {
            name: s["description"] for name, s in SCENARIOS.items()
        },
        "countries": list(COUNTRIES.keys()),
        "crop_types": CROP_TYPES,
        "insurance_types": INSURANCE_TYPES,
        "seasons": SEASONS,
        "sources": [
            "Jensen & Barrett (2016) AJAE - IBLI basis risk analysis",
            "WFP R4 Rural Resilience Initiative reports (2018-2023)",
            "ACRE Africa scaling data (GIIF 2016, IFC 2014)",
            "Ntukamazina et al. (2017) J. Agr. Rural Develop. Trop. Subtrop.",
            "IPCC AR6 WGII Chapter 9 - Africa climate projections",
        ],
        "variables": {
            "record_id": "Unique integer identifier",
            "country": "Country name",
            "year": "Calendar year (2015-2024)",
            "season": "Growing season (long_rains / short_rains)",
            "crop_type": "Insured crop (maize/wheat/sorghum/rice/cassava)",
            "insurance_type": "Product type (index_based/area_yield/weather_derivative)",
            "scenario": "Policy scenario (baseline/expanded_index_insurance/climate_shock)",
            "enrolled_hectares": "Hectares covered by the policy",
            "premium_paid_usd": "Premium amount in USD",
            "ndvi_anomaly": "NDVI z-score anomaly (negative = vegetation deficit)",
            "rainfall_deficit_pct": "Percentage below normal rainfall",
            "yield_loss_pct": "Estimated yield loss percentage",
            "payout_triggered": "Whether index triggered a payout (True/False)",
            "payout_amount_usd": "Payout disbursement in USD (0 if not triggered)",
            "payout_ratio": "Payout as fraction of insured value",
            "farmer_count": "Number of farmers in the micro-cohort",
            "enrollment_rate": "Regional enrollment rate (0-1)",
            "lapse_rate": "Policy non-renewal rate (0-1)",
            "basis_risk_score": "Residual unhedged risk (0=perfect, 1=none)",
        },
        "parameter_sources": {
            "premium_rates": "IBLI: 5.5%/3.25% of insured value; R4: 10-15% farmer contribution",
            "payout_trigger": "IBLI: NDVI < 15th percentile; ACRE: yield < 80% of average",
            "basis_risk": "Jensen et al. 2016: IBLI reduces covariate risk by 63%, residual 69%",
            "enrollment": "R4: ~180K farmers (2020); ACRE: ~400K farmers (2015)",
            "sum_insured": "R4: ~$98/farmer; ACRE: $650/acre for bundled seed products",
        },
    }

    with open(os.path.join(output_dir, "metadata.json"), "w") as f:
        json.dump(metadata, f, indent=2)


if __name__ == "__main__":
    output_dir = os.path.join(os.path.dirname(__file__), "data")
    os.makedirs(output_dir, exist_ok=True)

    print("Generating 30,000 records (12 countries × 3 scenarios × ~10,000)...")
    records = generate_dataset()

    csv_path = os.path.join(output_dir, "african_crop_insurance_payouts.csv")
    write_csv(records, csv_path)
    print(f"Wrote {len(records)} records to {csv_path}")

    write_metadata(os.path.dirname(__file__))
    print("Wrote metadata.json")

    # Print summary stats
    payouts = [r for r in records if r["payout_triggered"]]
    print(f"\nSummary:")
    print(f"  Total records: {len(records)}")
    print(f"  Payouts triggered: {len(payouts)} ({100*len(payouts)/len(records):.1f}%)")
    print(f"  Countries: {len(COUNTRIES)}")
    print(f"  Scenarios: {list(SCENARIOS.keys())}")
    if payouts:
        print(f"  Avg payout: ${np.mean([r['payout_amount_usd'] for r in payouts]):.2f}")
        print(f"  Avg basis risk: {np.mean([r['basis_risk_score'] for r in records]):.3f}")