A Hybrid Smoothing Model for Historical Reconstruction of Route-Level Airline Passenger Demand

Abstract

Historical pax demand estimation is important for understanding long-term demand dynamics and supporting airline planning, strategic decision-making, and transportation policy analysis. However, current decomposition approaches either estimate trends without explicitly modeling seasonality or jointly estimate trend and seasonal components without providing user control over trend smoothness. This paper proposes a Guerrero–Holt–Winters (GHW) hybrid smoothing model to estimate and reconstruct historical route-level airline pax demand. The GHW model combines Guerrero’s user-controlled trend estimation with the multiplicative seasonal structure of the Holt–Winters (HW) model to produce an interpretable decomposition of historical demand into trend, seasonal, and irregular components. The model is applied to quarterly pax demand data for ten representative U.S. domestic air routes covering the period 2015–2023. The results show that the proposed GHW model achieves lower in-sample reconstruction errors than the multiplicative HW model at lower degrees of trend smoothness within the evaluated range. Among the evaluated specifications, λ = 1, corresponding to a Guerrero smoothness index of S(1; 36) = 58.82%, produces the lowest in-sample reconstruction error. The historical reconstruction error, measured by the root mean squared error (RMSE), is reduced by approximately 54.4% relative to the multiplicative HW model. Similar improvements are observed under MAE, MAPE, and sMAPE, indicating that the reduction in reconstruction error at this smoothness level is consistent across the four error measures. An expanded benchmark comparison with HW, ETS state-space exponential smoothing, STL decomposition, and HP1600 further shows that the GHW model produces the lowest in-sample reconstruction errors across all four evaluation error metrics and all ten routes analyzed. A sensitivity analysis further demonstrates that reconstruction accuracy declines as the degree of trend smoothness increases from S(λ; 36) = 58.82% to 88.63%, whereas the estimated seasonal factors remain remarkably stable across the evaluated smoothness range. A COVID-19 robustness analysis further shows that reconstruction errors are substantially lower in the Pre-COVID period than in the Post-COVID period across all ten routes, indicating that reconstruction accuracy is sensitive to the period analyzed.