Overview
This project forecasts total passenger arrivals at Dublin Airport for the next 12 months (April 2026 – March 2027) using monthly data from the Central Statistics Office (CSO) Ireland API (table ASM01). Dublin Airport accounts for approximately 83% of all Irish air arrivals.
The dataset spans January 2010 to March 2026 (195 monthly observations). A key methodological challenge was the COVID-19 structural break (January 2020 – February 2022), which caused a near-complete collapse in passenger volumes for 26 months. Two experimental setups were designed to isolate its impact on forecast accuracy.
Statistical validation was conducted via ADF and KPSS tests (stationarity), Ljung-Box (white noise), and Shapiro-Wilk (normality) before any model training.
Experimental Design
Full series, COVID included
Used as a baseline to quantify how the pandemic distorts model learning. ADF test confirms stationarity (p=0.029). Models: Exponential Smoothing, ARIMA, ETS, LASSO, LR.
COVID excluded · Drift imputation
The 26-month COVID gap is replaced with drift interpolation — a linear bridge preserving seasonal structure. ADF non-stationary (p=0.95), confirming the clean seasonal trend. Models: Exp. Smoothing, ARIMA, Extra Trees, OMP, GBR.
Both setups use expanding window cross-validation (3 folds), a 12-month forecast horizon and 90% confidence intervals. Models are tuned via randomised grid search (30 candidates, 3 folds each) and combined into a blended ensemble (weighted average of the top 3).
Models Benchmarked
| Model | Type | Setup | Notes |
|---|---|---|---|
| Exponential Smoothing ★ Best B | Statistical | A + B | Best SMAPE in Setup B after hyperparameter tuning |
| ETS ★ Best A | Statistical | A | Best in Setup A; captures additive seasonality + trend |
| ARIMA | Statistical | A + B | Auto-order selection; tuned via grid search |
| Gradient Boosted Regression | ML (CDS+DT) | B | Highest peak forecast (~2,058k in Jul 2026) |
| Extra Trees | ML (CDS+DT) | B | Conservative forecast; stable interval |
| Blended Ensemble | Ensemble | A + B | Weighted average of top 3 from compare_models |
12-Month Forecast — Setup B (Primary)
Best model: Exponential Smoothing. Values in thousands of passengers. Strong seasonal pattern — peak in summer (Jul–Aug), trough in winter (Jan–Feb).
| Month | Exp. Smoothing | ARIMA | GBR | Blend |
|---|---|---|---|---|
| Apr 2026 | 1,561 | 1,566 | 1,596 | 1,558 |
| May 2026 | 1,678 | 1,705 | 1,730 | 1,664 |
| Jun 2026 | 1,773 | 1,768 | 1,850 | 1,749 |
| Jul 2026 ★ | 1,952 | 1,924 | 2,058 | 1,906 |
| Aug 2026 ★ | 1,966 | 1,972 | 2,086 | 1,926 |
| Sep 2026 | 1,721 | 1,709 | 1,851 | 1,712 |
| Oct 2026 | 1,638 | 1,641 | 1,768 | 1,618 |
| Nov 2026 | 1,346 | 1,326 | 1,480 | 1,378 |
| Dec 2026 | 1,316 | 1,323 | 1,464 | 1,351 |
| Jan 2027 | 1,257 | 1,244 | 1,373 | 1,304 |
| Feb 2027 | 1,249 | 1,200 | 1,383 | 1,302 |
| Mar 2027 | 1,403 | 1,450 | 1,671 | 1,438 |
★ Peak summer months. All Setup B models agree on Jul–Aug as the seasonal peak, with Exp. Smoothing and ARIMA converging around 1,950–1,970k passengers in August 2026.