Timing-Aware Container-Throughput Forecasting Under İnformation and Uncertainty Constraints

Abstract

Monthly container throughput in Türkiye is examined for one- and three-month forecast horizons using 172 observations from 2012M01 to 2026M04 and an expanding-window evaluation. For the national series, Ridge regularization is selected inside each outer training window, while component comparisons keep the Ridge penalty fixed across origins. Because the h= 1 forecast origin follows publication of the previous month’s UAB statistics, it occurs after the target month has begun and is therefore interpreted operationally as a current-month nowcast rather than a forecast issued before the target month. Among the national specifications, the level-target operational-history Ridge gives the lowest forecast error. Compared with seasonal naive, mean MAE is lower by 31.2% at h= 1 and 27.0% at h= 3. These gains fall to about 20% after excluding February–March 2024, and the Harvey–Leybourne–Newbold-corrected tests do not reach the 5% level. Calendar-lagged trade does not lower national MAE. At h= 1, MAE is 1.9% higher even though the Clark–West test is significant (p = 0.027). At h= 3, MAE is 11.1% higher and p = 0.845. None of the 28 component-level lagged-trade additions is FDR-supported. In the retrospective same-period experiment, contemporaneous trade lowers loading-total MAE by 6.0% (p = 0.004; q = 0.027), but these values are unavailable at the forecast origin. For seasonal-log-difference targets, only Elastic Net improves on seasonal naive at both horizons, while all transformed-target models remain less accurate than the level-target Ridge. Sequential conformal prediction gives 82.1% coverage at h= 1 and 89.3% at h= 3 for the 80% nominal level.