Forecasting monthly global IC export values for the remainder of 2025 using classical and regression-based time-series models. We compare SNaive12, Linear Trend with Seasonality (LTwS/CMAT), Regression with Fourier seasonality, and a Hybrid (Regression + lag residuals) to select a production forecast for planning/S&OP.
Result: Strong seasonality (Feb low; Sep–Nov high). Hybrid wins out-of-sample (≈ $0.13T RMSE). 6-month outlook: moderate rebound with seasonal peaks in Aug–Oct.
| Trend & Seasonality | Backtest (Walk-forward) | 2025 Forecast |
|---|---|---|
 |
 |
 |
Reading notes (from slides):
- Post-COVID rebound, 2020→2022 surge, then stabilization.
- Baseline SNaive12 performs decently due to strong seasonality, but Hybrid improves RMSE and realism.
Data: UN Comtrade monthly exports (Jan-2020 → Jun-2025).
Filters: exports only; drop aggregates (“World”), positive values; prefer primaryValue (fallback fobvalue); month index built from period (YYYY-MM).
Models evaluated:
- SNaive12 (seasonal naïve, same month last year)
- LTwS/CMAT (linear trend on centered moving average + seasonal indices)
- Regression w/ seasonality: time trend + month dummies / Fourier (K=1..2)
- Hybrid: Regression + lagged residual model (adds short-memory correction)
Evaluation:
- Walk-forward / expanding window backtest
- Primary metric: RMSE (lower is better)
- Parameter sweep for Ridge α, Fourier K, lag set, and min train window.
- Seasonality dominates: SNaive12 is a strong baseline → confirms monthly pattern stability.
- Regression + Fourier captures smooth seasonality; LTwS is conservative.
- Hybrid reliably reduces forecast error by modeling residual short-memory components.
- Operational takeaways (slides): Plan pull-up for Sep–Nov; anticipate Feb slowdown; use Hybrid as central forecast with ±$0.13T band for budgeting/S&OP.
- Project slides: “Final Project DS IC Value Forecasting on 2025.pdf.” Figures and summary text above are sourced from these slides.
- Data: UN Comtrade
- Author: Rizky Febri Ibra Habibie
- Contact: rizkyfebriibrahabibie@gmail.com · LinkedIn: /in/rizkyfebriibrahabibie/