이상치 탐지를 SQL로: Z-Score와 IQR 기반 아웃라이어 찾기

비정상 데이터를 SQL로 자동 탐지. Z-Score, IQR, 백분위 기반 이상치 탐지 구현.

TL;DR

이상치(Outlier): 정상 범위를 벗어난 데이터 포인트
Z-Score: 평균에서 몇 표준편차 떨어졌는지 (|Z| > 3이면 이상치)
IQR: 사분위 범위 기반 (Q1-1.5×IQR ~ Q3+1.5×IQR 벗어나면 이상치)
활용: 이상 거래, 비정상 트래픽, 데이터 오류 탐지

1. 왜 이상치 탐지가 필요한가?

실제 사례

sql

일별 매출: 100만, 105만, 98만, 102만, 950만(??), 101만

950만원은 정상인가?

시스템 오류?
대량 주문?
사기 거래?

→ 자동으로 감지해서 알림 필요!

이상치 유형

유형	예시
데이터 오류	음수 주문금액, NULL 값
이상 거래	갑자기 10배 큰 주문
사기 탐지	비정상 패턴의 로그인
시스템 이상	트래픽 급증/급락

2. Z-Score 방식

기본 개념

sql

Z = (x - μ) / σ

x: 개별 값
μ: 평균
σ: 표준편차

Z

SQL 구현

sql

WITH stats AS (
    SELECT
        AVG(amount) as mean_amount,
        STDDEV(amount) as std_amount
    FROM orders
    WHERE created_at >= CURRENT_DATE - INTERVAL '30 days'
),
z_scores AS (
    SELECT
        o.order_id,
        o.user_id,
        o.amount,
        o.created_at,
        (o.amount - s.mean_amount) / NULLIF(s.std_amount, 0) as z_score
    FROM orders o
    CROSS JOIN stats s
    WHERE o.created_at >= CURRENT_DATE - INTERVAL '30 days'
)
SELECT
    order_id,
    user_id,
    amount,
    created_at,
    ROUND(z_score, 2) as z_score,
    CASE
        WHEN z_score > 3 THEN 'HIGH_OUTLIER'
        WHEN z_score < -3 THEN 'LOW_OUTLIER'
        ELSE 'NORMAL'
    END as status
FROM z_scores
WHERE ABS(z_score) > 3
ORDER BY ABS(z_score) DESC;

결과:

order_id	user_id	amount	z_score	status
12345	789	9500000	8.52	HIGH_OUTLIER
12346	790	5000	-3.21	LOW_OUTLIER

3. IQR (사분위 범위) 방식

기본 개념

sql

IQR = Q3 - Q1
하한 = Q1 - 1.5 × IQR
상한 = Q3 + 1.5 × IQR

SQL 구현

sql

WITH quartiles AS (
    SELECT
        PERCENTILE_CONT(0.25) WITHIN GROUP (ORDER BY amount) as q1,
        PERCENTILE_CONT(0.50) WITHIN GROUP (ORDER BY amount) as q2,
        PERCENTILE_CONT(0.75) WITHIN GROUP (ORDER BY amount) as q3
    FROM orders
    WHERE created_at >= CURRENT_DATE - INTERVAL '30 days'
),
bounds AS (
    SELECT
        q1,
        q2,
        q3,
        q3 - q1 as iqr,
        q1 - 1.5 * (q3 - q1) as lower_bound,
        q3 + 1.5 * (q3 - q1) as upper_bound
    FROM quartiles
)
SELECT
    o.order_id,
    o.user_id,
    o.amount,
    o.created_at,
    b.lower_bound,
    b.upper_bound,
    CASE
        WHEN o.amount > b.upper_bound THEN 'HIGH_OUTLIER'
        WHEN o.amount < b.lower_bound THEN 'LOW_OUTLIER'
        ELSE 'NORMAL'
    END as status
FROM orders o
CROSS JOIN bounds b
WHERE o.created_at >= CURRENT_DATE - INTERVAL '30 days'
  AND (o.amount > b.upper_bound OR o.amount < b.lower_bound)
ORDER BY o.amount DESC;

4. Modified Z-Score (MAD 기반)

일반 Z-Score의 문제

극단적 이상치에 평균/표준편차가 왜곡됨
해결: 중앙값(Median)과 MAD 사용

SQL 구현

sql

WITH median_calc AS (
    SELECT
        PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY amount) as median_amount
    FROM orders
    WHERE created_at >= CURRENT_DATE - INTERVAL '30 days'
),
mad_calc AS (
    SELECT
        m.median_amount,
        PERCENTILE_CONT(0.5) WITHIN GROUP (
            ORDER BY ABS(o.amount - m.median_amount)
        ) as mad
    FROM orders o
    CROSS JOIN median_calc m
    WHERE o.created_at >= CURRENT_DATE - INTERVAL '30 days'
),
modified_z AS (
    SELECT
        o.order_id,
        o.amount,
        m.median_amount,
        m.mad,
        0.6745 * (o.amount - m.median_amount) / NULLIF(m.mad, 0) as modified_z_score
    FROM orders o
    CROSS JOIN mad_calc m
    WHERE o.created_at >= CURRENT_DATE - INTERVAL '30 days'
)
SELECT
    order_id,
    amount,
    ROUND(modified_z_score, 2) as modified_z,
    CASE
        WHEN ABS(modified_z_score) > 3.5 THEN 'OUTLIER'
        ELSE 'NORMAL'
    END as status
FROM modified_z
WHERE ABS(modified_z_score) > 3.5
ORDER BY ABS(modified_z_score) DESC;

5. 시계열 이상치 탐지

이동평균 기반 이상치

sql

WITH daily_metrics AS (
    SELECT
        DATE(created_at) as date,
        SUM(amount) as daily_revenue,
        COUNT(*) as daily_orders
    FROM orders
    GROUP BY DATE(created_at)
),
with_moving_stats AS (
    SELECT
        date,
        daily_revenue,
        daily_orders,
        AVG(daily_revenue) OVER (
            ORDER BY date
            ROWS BETWEEN 7 PRECEDING AND 1 PRECEDING
        ) as ma_7d,
        STDDEV(daily_revenue) OVER (
            ORDER BY date
            ROWS BETWEEN 7 PRECEDING AND 1 PRECEDING
        ) as std_7d
    FROM daily_metrics
)
SELECT
    date,
    daily_revenue,
    ROUND(ma_7d, 0) as ma_7d,
    ROUND(std_7d, 0) as std_7d,
    ROUND((daily_revenue - ma_7d) / NULLIF(std_7d, 0), 2) as z_score,
    CASE
        WHEN (daily_revenue - ma_7d) / NULLIF(std_7d, 0) > 2 THEN 'SPIKE'
        WHEN (daily_revenue - ma_7d) / NULLIF(std_7d, 0) < -2 THEN 'DROP'
        ELSE 'NORMAL'
    END as status
FROM with_moving_stats
WHERE ma_7d IS NOT NULL
ORDER BY date DESC;

6. 사용자별 이상 행동 탐지

개인 패턴 대비 이상치

sql

WITH user_stats AS (
    SELECT
        user_id,
        AVG(amount) as user_avg,
        STDDEV(amount) as user_std,
        COUNT(*) as order_count
    FROM orders
    WHERE created_at >= CURRENT_DATE - INTERVAL '90 days'
    GROUP BY user_id
    HAVING COUNT(*) >= 5  -- 최소 5건 이상
),
recent_orders AS (
    SELECT
        o.order_id,
        o.user_id,
        o.amount,
        o.created_at,
        us.user_avg,
        us.user_std,
        (o.amount - us.user_avg) / NULLIF(us.user_std, 0) as user_z_score
    FROM orders o
    JOIN user_stats us ON o.user_id = us.user_id
    WHERE o.created_at >= CURRENT_DATE - INTERVAL '7 days'
)
SELECT
    order_id,
    user_id,
    amount,
    ROUND(user_avg, 0) as user_avg,
    ROUND(user_z_score, 2) as z_vs_user_pattern,
    CASE
        WHEN user_z_score > 3 THEN 'UNUSUAL_HIGH'
        WHEN user_z_score < -2 THEN 'UNUSUAL_LOW'
        ELSE 'NORMAL'
    END as anomaly_flag
FROM recent_orders
WHERE ABS(user_z_score) > 2
ORDER BY ABS(user_z_score) DESC;

7. 다중 지표 이상치

복합 조건으로 이상 탐지

sql

WITH order_features AS (
    SELECT
        order_id,
        user_id,
        amount,
        items_count,
        created_at,
        -- 시간대
        EXTRACT(HOUR FROM created_at) as order_hour,
        -- 요일
        EXTRACT(DOW FROM created_at) as order_dow
    FROM orders
    WHERE created_at >= CURRENT_DATE - INTERVAL '30 days'
),
global_stats AS (
    SELECT
        AVG(amount) as avg_amount,
        STDDEV(amount) as std_amount,
        AVG(items_count) as avg_items,
        STDDEV(items_count) as std_items
    FROM order_features
),
anomaly_scores AS (
    SELECT
        of.order_id,
        of.user_id,
        of.amount,
        of.items_count,
        of.created_at,
        (of.amount - gs.avg_amount) / NULLIF(gs.std_amount, 0) as z_amount,
        (of.items_count - gs.avg_items) / NULLIF(gs.std_items, 0) as z_items,
        -- 새벽 시간 주문
        CASE WHEN of.order_hour BETWEEN 2 AND 5 THEN 1 ELSE 0 END as late_night_flag
    FROM order_features of
    CROSS JOIN global_stats gs
)
SELECT
    order_id,
    user_id,
    amount,
    items_count,
    created_at,
    ROUND(z_amount, 2) as z_amount,
    ROUND(z_items, 2) as z_items,
    late_night_flag,
    -- 복합 이상 점수
    ROUND(ABS(z_amount) + ABS(z_items) + late_night_flag * 2, 2) as anomaly_score
FROM anomaly_scores
WHERE ABS(z_amount) > 2 OR ABS(z_items) > 2 OR late_night_flag = 1
ORDER BY anomaly_score DESC
LIMIT 100;

8. 백분위 기반 탐지

상위/하위 N% 탐지

sql

WITH percentile_bounds AS (
    SELECT
        PERCENTILE_CONT(0.01) WITHIN GROUP (ORDER BY amount) as p1,
        PERCENTILE_CONT(0.99) WITHIN GROUP (ORDER BY amount) as p99
    FROM orders
    WHERE created_at >= CURRENT_DATE - INTERVAL '30 days'
)
SELECT
    o.order_id,
    o.user_id,
    o.amount,
    o.created_at,
    CASE
        WHEN o.amount >= pb.p99 THEN 'TOP_1%'
        WHEN o.amount <= pb.p1 THEN 'BOTTOM_1%'
        ELSE 'NORMAL'
    END as percentile_status,
    ROUND(PERCENT_RANK() OVER (ORDER BY o.amount) * 100, 2) as percentile
FROM orders o
CROSS JOIN percentile_bounds pb
WHERE o.created_at >= CURRENT_DATE - INTERVAL '7 days'
  AND (o.amount >= pb.p99 OR o.amount <= pb.p1)
ORDER BY o.amount DESC;

9. 실시간 알림 쿼리

즉시 감지해야 할 이상치

sql

WITH recent_baseline AS (
    SELECT
        AVG(amount) as avg_amount,
        STDDEV(amount) as std_amount,
        PERCENTILE_CONT(0.99) WITHIN GROUP (ORDER BY amount) as p99
    FROM orders
    WHERE created_at >= CURRENT_DATE - INTERVAL '7 days'
      AND created_at < CURRENT_TIMESTAMP - INTERVAL '1 hour'
),
last_hour_orders AS (
    SELECT
        order_id,
        user_id,
        amount,
        created_at
    FROM orders
    WHERE created_at >= CURRENT_TIMESTAMP - INTERVAL '1 hour'
)
SELECT
    o.order_id,
    o.user_id,
    o.amount,
    o.created_at,
    ROUND((o.amount - b.avg_amount) / NULLIF(b.std_amount, 0), 2) as z_score,
    CASE
        WHEN o.amount > b.p99 * 2 THEN 'CRITICAL'
        WHEN o.amount > b.p99 THEN 'WARNING'
        WHEN (o.amount - b.avg_amount) / NULLIF(b.std_amount, 0) > 3 THEN 'WARNING'
        ELSE 'OK'
    END as alert_level
FROM last_hour_orders o
CROSS JOIN recent_baseline b
WHERE o.amount > b.p99
   OR (o.amount - b.avg_amount) / NULLIF(b.std_amount, 0) > 3
ORDER BY o.amount DESC;

10. 이상치 요약 대시보드

일별 이상치 현황

sql

WITH daily_orders AS (
    SELECT
        DATE(created_at) as date,
        order_id,
        amount
    FROM orders
    WHERE created_at >= CURRENT_DATE - INTERVAL '30 days'
),
daily_stats AS (
    SELECT
        date,
        COUNT(*) as total_orders,
        AVG(amount) as avg_amount,
        STDDEV(amount) as std_amount
    FROM daily_orders
    GROUP BY date
),
outlier_counts AS (
    SELECT
        DATE(o.created_at) as date,
        COUNT(*) as outlier_count,
        SUM(o.amount) as outlier_amount
    FROM orders o
    JOIN daily_stats ds ON DATE(o.created_at) = ds.date
    WHERE (o.amount - ds.avg_amount) / NULLIF(ds.std_amount, 0) > 3
      AND o.created_at >= CURRENT_DATE - INTERVAL '30 days'
    GROUP BY DATE(o.created_at)
)
SELECT
    ds.date,
    ds.total_orders,
    COALESCE(oc.outlier_count, 0) as outliers,
    ROUND(COALESCE(oc.outlier_count, 0) * 100.0 / ds.total_orders, 2) as outlier_pct,
    ROUND(ds.avg_amount, 0) as avg_amount,
    COALESCE(ROUND(oc.outlier_amount, 0), 0) as outlier_amount
FROM daily_stats ds
LEFT JOIN outlier_counts oc ON ds.date = oc.date
ORDER BY ds.date DESC;

결론

방법	장점	단점	사용 시기
Z-Score	간단, 직관적	극단값에 민감	정규분포 데이터
IQR	극단값에 강건	분포 가정 없음	비대칭 데이터
Modified Z	강건함	계산 복잡	극단값 많을 때
이동평균	시계열 적합	지연 발생	일별/시간별 모니터링
백분위	이해 쉬움	상대적 기준	Top/Bottom N%

이상치 탐지 핵심:

단일 방법보다 복합 조건 사용
전역 통계 + 개인 패턴 비교
실시간 알림 체계 구축

References

NIST - Guidelines for Outlier Detection
Scikit-learn - Outlier Detection Methods
PostgreSQL - Statistical Functions