Materi 10 · Advanced Techniques

Forecasting & Time Series

Prediksi masa depan dari data masa lalu. Trend, seasonality, ARIMA, Prophet, dan exponential smoothing — toolkit untuk demand planning, financial forecasting, dan workforce planning.

⏱ 28 Menit🎯 Intermediate📅 Forecasting

1. Anatomi Time Series

4 Komponen Utama

Trend: arah jangka panjang (naik / turun / flat).
Seasonality: pola berulang dalam periode tetap (mingguan, bulanan, tahunan).
Cyclical: fluktuasi jangka panjang non-fixed (siklus ekonomi).
Irregular / Noise: random variation yang tidak bisa diprediksi.

2. Decomposition

Additive Model Y = Trend + Seasonality + Residual
// Cocok kalau seasonal effect konstan

Multiplicative Model Y = Trend × Seasonality × Residual
// Cocok kalau seasonal effect proporsional dengan magnitude

Contoh: Penjualan ritel seasonal multiplicative — Lebaran lift 30% (proportional). Suhu harian additive — winter -20°C tetap konstan.

3. Metode Forecasting Klasik

Metode	Karakter	Kapan Pakai
Naive	Ŷ_t+1 = Y_t	Baseline benchmark
Moving Average	Rata-rata N period terakhir	Smooth noise, identify trend
Exponential Smoothing	Recent data weighted higher	Stable data tanpa seasonality
Holt's Linear	ES + trend	Data dengan trend tapi tanpa seasonality
Holt-Winters	ES + trend + seasonality	Data dengan semua komponen
ARIMA	Auto-regressive	Series stationary
SARIMA	ARIMA + seasonal	Series dengan seasonality
Prophet (Meta)	Decomposable, robust	Business series, mudah dipakai

4. STATIONARITY

Wajib Stationary untuk ARIMA

Time series stationary = mean & variance konstan dari waktu ke waktu, tanpa trend/seasonality. ARIMA butuh ini. Jika tidak stationary: differencing (Y_t - Y_t-1) untuk hilangkan trend.

ADF Test (Augmented Dickey-Fuller): test stationarity. p-value < 0.05 → stationary.
Differencing: first-order untuk hilang trend, seasonal differencing untuk hilang seasonality.
Log transform: stabilize variance untuk skewed data.
Detrending: remove fitted trend line.

5. ARIMA — Auto Regressive Integrated Moving Average

ARIMA(p, d, q) p = order of AR (lag value)
d = degree of differencing
q = order of MA (lag error)
// Pakai ACF & PACF plot untuk identify p & q

SARIMA(p,d,q)(P,D,Q,m) (p,d,q): non-seasonal terms
(P,D,Q): seasonal terms
m: seasonal period (12 untuk monthly, 7 untuk weekly)

6. PROPHET — Modern Easy

Meta's Prophet — Robust & User-Friendly

Open source library dari Meta. Handle missing data, outliers, holidays, multiple seasonality automatically. Cukup butuh kolom 'ds' (date) dan 'y' (value). Best balance antara accuracy dan ease-of-use.

Prophet di Python from prophet import Prophet
m = Prophet()
m.fit(df)
future = m.make_future_dataframe(periods=30)
forecast = m.predict(future)
m.plot(forecast)

7. EVALUATION METRICS

Forecast Accuracy MAE = (1/n) × Σ|y - ŷ| // intuitive
MAPE = (1/n) × Σ|y - ŷ|/y × 100% // percentage error
RMSE = √((1/n) × Σ(y - ŷ)²) // penalize big errors
MASE = MAE / MAE_naive // vs naive baseline

Selalu Pakai Baseline

Model ARIMA fancy tapi MAE-nya 1.2× naive forecast = useless. Naive forecast (Y_t+1 = Y_t) atau seasonal naive harus jadi floor benchmark. Model harus beat ini significantly.

8. Backtesting & Walk-Forward Validation

Train-test split: jangan random — pakai chronological. Train: 80% awal, test: 20% akhir.
Walk-forward validation: retrain pada window yang bergeser. Lebih realistis.
Holdout multiple periods: evaluate di 3-6 bulan terakhir terpisah.
Track residuals: residual harus random, tidak ada pola = model bagus.

9. Studi Kasus

Demand Forecasting di E-Commerce

E-commerce besar pakai Prophet + holiday effect untuk forecast inventory. Mereka encode Lebaran, Black Friday, gaji-an (tanggal 25-akhir bulan) sebagai event. Hasilnya: stockout berkurang 38%, overstock berkurang 22%. Pelajaran: domain knowledge mengalahkan model fancy.

📝 Tugas Praktik

Cari dataset time series di Kaggle (mis. retail sales monthly).
Plot data. Identifikasi trend, seasonality, outlier secara visual.
Decompose data dengan statsmodels Python.
Train 3 model: Naive baseline, Holt-Winters, Prophet.
Hitung MAE, MAPE, RMSE pada test set. Bandingkan vs baseline.
Forecast 12 period ke depan + confidence interval.

Rangkuman

Time series = Trend + Seasonality + Cyclical + Noise.
Decomposition: Additive (constant) vs Multiplicative (proportional).
ARIMA butuh stationarity. Differencing untuk transform.
Prophet (Meta) = balance accuracy + ease-of-use untuk business series.
MAPE intuitive, RMSE penalize big error, MASE bandingkan vs naive.
Selalu beat naive baseline. Backtest dengan walk-forward.
Domain knowledge (event, holiday) sering kalahkan model fancy.