2nd/10_Wiki/Topics/AI_and_ML/Predictive-Analytics.md

---
id: wiki-2026-0508-predictive-analytics
title: Predictive Analytics
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Predictive Modeling, Forecasting Analytics]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [ml, analytics, forecasting, regression, time-series]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: scikit-learn / XGBoost / Prophet / PyTorch Forecasting
---

# Predictive Analytics

## 매 한 줄
> **"매 historical data 로 future outcomes 예측 — regression, classification, time series 의 union"**. 1990s statistics 에서 출발해 2010s ML 으로 mainstream, 2026 currently transformer-based forecasting (TimesFM, Chronos) 이 tabular 와 sequence 에서 공존. Business intelligence, supply chain, churn, fraud 의 매 backbone.

## 매 핵심

### 매 problem types
- **Regression**: continuous target (revenue, demand, price).
- **Classification**: discrete label (churn yes/no, fraud type).
- **Time series**: temporally indexed (sales, sensor, stock).
- **Survival**: time-to-event (customer lifetime, equipment failure).
- **Ranking**: ordering items (recommendation, search).

### 매 modern stack (2026)
- **Tabular**: XGBoost / LightGBM / CatBoost — 매 still SOTA on most tabular.
- **Deep tabular**: TabPFN v2, FT-Transformer — 매 zero-shot tabular foundation.
- **Time series**: Chronos (Amazon), TimesFM (Google), Moirai — 매 pretrained TS foundation models.
- **Classic TS**: Prophet, statsforecast (AutoARIMA, ETS) — 매 baseline.

### 매 workflow
1. EDA + feature engineering.
2. Train/val/test split (temporal for TS).
3. Model selection + CV.
4. Hyperparameter tuning (Optuna).
5. Calibration + interpretability (SHAP).
6. Deployment + monitoring (drift detection).

## 💻 패턴

### XGBoost regression baseline
```python
import xgboost as xgb
from sklearn.model_selection import KFold
from sklearn.metrics import mean_absolute_error
import numpy as np

X, y = load_data()
kf = KFold(n_splits=5, shuffle=True, random_state=42)
scores = []
for tr, va in kf.split(X):
    model = xgb.XGBRegressor(
        n_estimators=2000, learning_rate=0.03, max_depth=6,
        subsample=0.8, colsample_bytree=0.8,
        early_stopping_rounds=50, eval_metric="mae",
    )
    model.fit(X.iloc[tr], y.iloc[tr], eval_set=[(X.iloc[va], y.iloc[va])], verbose=False)
    scores.append(mean_absolute_error(y.iloc[va], model.predict(X.iloc[va])))
print(f"CV MAE: {np.mean(scores):.4f} +/- {np.std(scores):.4f}")
```

### LightGBM classification w/ early stopping
```python
import lightgbm as lgb
model = lgb.LGBMClassifier(
    n_estimators=5000, learning_rate=0.02, num_leaves=63,
    min_child_samples=20, reg_alpha=0.1, reg_lambda=0.1,
)
model.fit(
    X_tr, y_tr,
    eval_set=[(X_va, y_va)],
    callbacks=[lgb.early_stopping(100), lgb.log_evaluation(0)],
)
proba = model.predict_proba(X_te)[:, 1]
```

### Time series with TimesFM (foundation model, zero-shot)
```python
import timesfm
tfm = timesfm.TimesFm(
    hparams=timesfm.TimesFmHparams(backend="gpu", per_core_batch_size=32),
    checkpoint=timesfm.TimesFmCheckpoint(huggingface_repo_id="google/timesfm-2.0-500m"),
)
forecast, _ = tfm.forecast(
    inputs=[history_series],     # list of np.array
    freq=[0],                     # 0=high freq, 1=med, 2=low
    horizon_len=96,
)
```

### Prophet (interpretable seasonality)
```python
from prophet import Prophet
m = Prophet(yearly_seasonality=True, weekly_seasonality=True, changepoint_prior_scale=0.05)
m.add_country_holidays(country_name="US")
m.fit(df)  # df with ds, y columns
future = m.make_future_dataframe(periods=90)
fcst = m.predict(future)  # yhat, yhat_lower, yhat_upper
```

### SHAP explainability
```python
import shap
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_te)
shap.summary_plot(shap_values, X_te)
shap.waterfall_plot(shap.Explanation(values=shap_values[0], base_values=explainer.expected_value, data=X_te.iloc[0]))
```

### Probability calibration
```python
from sklearn.calibration import CalibratedClassifierCV
calibrated = CalibratedClassifierCV(base_model, method="isotonic", cv="prefit")
calibrated.fit(X_va, y_va)
# Brier score + reliability diagram on test
```

### Drift monitoring (Evidently)
```python
from evidently.report import Report
from evidently.metric_preset import DataDriftPreset
report = Report(metrics=[DataDriftPreset()])
report.run(reference_data=X_train, current_data=X_prod)
report.save_html("drift.html")
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Tabular, <1M rows | XGBoost / LightGBM |
| Tabular, mixed features | CatBoost |
| Zero-shot tabular | TabPFN v2 |
| Time series, single series | Prophet / AutoARIMA |
| Time series, many series, zero-shot | Chronos / TimesFM |
| Need interpretability | Linear / GAM + SHAP |
| Very large (>10M) tabular | LightGBM w/ histogram |

**기본값**: LightGBM + Optuna + SHAP for tabular; Chronos zero-shot for new TS problems.

## 🔗 Graph
- 부모: [[Machine-Learning]] · [[Statistics]]
- 변형: [[Time-Series-Analysis|Time-Series-Forecasting]] · [[Regression]]
- Adjacent: [[XGBoost]] · [[LightGBM]] · [[SHAP]] · [[Prophet]]

## 🤖 LLM 활용
**언제**: business outcome forecasting, risk scoring, demand planning, anomaly detection 의 supervised learning.
**언제 X**: causal inference (use [[Causal-Inference]]), prescriptive optimization (use [[Optimization]]).

## ❌ 안티패턴
- **Data leakage**: 매 future info in training features — invalidates evaluation.
- **Random split on time series**: must use temporal split.
- **Ignoring calibration**: probabilities used for decisions but never calibrated.
- **No drift monitoring**: model decays silently in production.
- **Over-engineering deep nets for small tabular**: GBDT wins under 100k rows.

## 🧪 검증 / 중복
- Verified (Kaggle Grandmaster patterns, Amazon Chronos paper 2024, Google TimesFM 2024).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full predictive analytics workflow with 2026 foundation models |