import numpy as np
import pandas as pd
from datetime import datetime, timedelta

np.random.seed(42)

n = 10000

age_groups = ["18-24", "25-34", "35-44", "45-54", "55+"]
regions = ["North", "South", "East", "West", "Central"]
income_bands = ["Low", "Medium", "High"]
channels = ["Email", "SMS", "App Push", "Paid Social"]
campaigns = [
    "New Year Retention",
    "Winter Clearance",
    "Loyalty Boost",
    "Spring Bundle",
    "Midyear Discount",
    "Premium Upgrade"
]

data = []

for i in range(1, n + 1):
    age_group = np.random.choice(
        age_groups,
        p=[0.16, 0.25, 0.25, 0.20, 0.14]
    )

    region = np.random.choice(regions)
    income_band = np.random.choice(
        income_bands,
        p=[0.38, 0.42, 0.20]
    )

    campaign_name = np.random.choice(campaigns)
    campaign_channel = np.random.choice(
        channels,
        p=[0.36, 0.18, 0.30, 0.16]
    )

    previous_purchase = np.random.binomial(1, 0.72)
    purchase_frequency_12m = np.random.poisson(2.8)
    avg_order_value = round(np.random.uniform(18, 95), 2)
    total_spend_12m = round(avg_order_value * purchase_frequency_12m, 2)

    historical_open_rate = round(np.random.uniform(0.10, 0.55), 3)
    historical_click_rate = round(np.random.uniform(0.02, 0.28), 3)

    engagement_score = round(
        (historical_open_rate * 0.45)
        + (historical_click_rate * 0.35)
        + (previous_purchase * 0.20),
        3
    )

    # Age and income effects are intentionally included
    # to create realistic fairness monitoring patterns.
    age_effect = {
        "18-24": 0.02,
        "25-34": 0.04,
        "35-44": 0.05,
        "45-54": 0.02,
        "55+": -0.08
    }[age_group]

    income_effect = {
        "Low": -0.03,
        "Medium": 0.02,
        "High": 0.05
    }[income_band]

    channel_effect = {
        "Email": 0.04,
        "App Push": 0.03,
        "Paid Social": 0.02,
        "SMS": -0.01
    }[campaign_channel]

    predicted_score = round(
        np.clip(
            0.45
            + engagement_score * 0.35
            + age_effect
            + income_effect
            + channel_effect
            + np.random.normal(0, 0.07),
            0.01,
            0.99
        ),
        3
    )

    targeting_threshold = 0.60
    model_recommendation = int(predicted_score >= targeting_threshold)

    # Not every recommended customer is targeted.
    # This simulates campaign capacity, control groups, and eligibility rules.
    marketing_eligible = np.random.binomial(1, 0.90)
    control_group = np.random.binomial(1, 0.15)

    targeted_flag = int(
        model_recommendation == 1
        and marketing_eligible == 1
        and control_group == 0
        and np.random.random() < 0.78
    )

    base_response_probability = (
        0.08
        + previous_purchase * 0.05
        + engagement_score * 0.18
        + targeted_flag * 0.12
        + channel_effect
        + np.random.normal(0, 0.035)
    )

    # Older customers are intentionally modeled with response potential
    # that the AI model sometimes underestimates.
    if age_group == "55+":
        base_response_probability += 0.035

    actual_response_probability = round(
        np.clip(base_response_probability, 0.01, 0.65),
        3
    )

    actual_response = np.random.binomial(1, actual_response_probability)
    conversion = actual_response

    revenue = round(conversion * np.random.uniform(22, 140), 2)
    gross_margin_pct = round(np.random.uniform(0.15, 0.52), 3)

    retention_30d = np.random.binomial(
        1,
        np.clip(0.56 + conversion * 0.10, 0, 1)
    )

    retention_90d = np.random.binomial(
        1,
        np.clip(0.45 + conversion * 0.08, 0, 1)
    )

    complaint_probability = 0.012
    opt_out_probability = 0.018

    if income_band == "Low":
        complaint_probability += 0.014
        opt_out_probability += 0.018

    if campaign_channel == "SMS":
        complaint_probability += 0.008
        opt_out_probability += 0.012

    complaint_flag = np.random.binomial(1, complaint_probability)
    opt_out_flag = np.random.binomial(1, opt_out_probability)
    unsubscribe_flag = np.random.binomial(1, opt_out_probability * 0.50)

    false_positive_flag = int(model_recommendation == 1 and actual_response == 0)
    false_negative_flag = int(model_recommendation == 0 and actual_response == 1)

    if false_negative_flag == 1 and age_group == "55+":
        risk_band = np.random.choice(["High", "Critical"], p=[0.78, 0.22])
    elif complaint_flag == 1 or opt_out_flag == 1:
        risk_band = np.random.choice(["Medium", "High"], p=[0.72, 0.28])
    else:
        risk_band = np.random.choice(["Low", "Medium"], p=[0.78, 0.22])

    campaign_date = datetime(2025, 1, 1) + timedelta(days=np.random.randint(0, 181))
    campaign_month = campaign_date.replace(day=1)

    data.append({
        "interaction_id": f"INT_{i:06d}",
        "customer_id": f"CUST_{np.random.randint(1, 4500):05d}",
        "campaign_id": f"CAMP_2025_{campaign_date.month:02d}",
        "campaign_name": campaign_name,
        "campaign_date": campaign_date.strftime("%Y-%m-%d"),
        "campaign_month": campaign_month.strftime("%Y-%m-%d"),
        "age_group": age_group,
        "region": region,
        "income_band": income_band,
        "campaign_channel": campaign_channel,
        "previous_purchase": previous_purchase,
        "purchase_frequency_12m": purchase_frequency_12m,
        "total_spend_12m": total_spend_12m,
        "avg_order_value": avg_order_value,
        "historical_open_rate": historical_open_rate,
        "historical_click_rate": historical_click_rate,
        "engagement_score": engagement_score,
        "model_version": "MKT_AI_v2.3",
        "targeting_threshold": targeting_threshold,
        "predicted_score": predicted_score,
        "model_recommendation": model_recommendation,
        "targeted_flag": targeted_flag,
        "control_group": control_group,
        "actual_response_probability": actual_response_probability,
        "actual_response": actual_response,
        "conversion": conversion,
        "revenue": revenue,
        "gross_margin_pct": gross_margin_pct,
        "retention_30d": retention_30d,
        "retention_90d": retention_90d,
        "complaint_flag": complaint_flag,
        "opt_out_flag": opt_out_flag,
        "unsubscribe_flag": unsubscribe_flag,
        "false_positive_flag": false_positive_flag,
        "false_negative_flag": false_negative_flag,
        "risk_band": risk_band
    })

df = pd.DataFrame(data)

df.to_csv(
    "ethical_marketing_ai_campaign_dataset_10000_PowerBI_SAFE_v2.csv",
    sep=";",
    index=False,
    encoding="utf-8-sig"
)