Hybrid ARIMA-LSTM Knowledge Management for Strategic Tourism Demand Forecasting in Morocco

Abstract

Strategic foresight represents one of the most consequential yet undertheorized capabilities in the knowledge management literature. Whilst organizations increasingly invest in AI-driven forecasting tools, these systems are predominantly designed and evaluated as technical artefacts rather than as knowledge management (KM) instruments. In doing so, they overlook the governance decisions, knowledge architectures, and organizational learning principles that determine whether such tools genuinely enhance an organization's capacity to anticipate and respond to environmental complexity. Tourism management epitomizes this challenge acutely. The sector's inherent complexity, pronounced seasonality, and vulnerability to exogenous shocks, as the COVID-19 pandemic dramatically illustrated, require forecasting approaches that function as organizational knowledge systems rather than mere statistical exercises. This paper addresses this gap by developing and evaluating a hybrid ARIMA-LSTM model for predicting monthly international tourist arrivals in Morocco, framed explicitly within the KM literature on knowledge creation, codification, and strategic foresight. Drawing on a 20-year multivariate dataset (2005–2024) that integrates macroeconomic, climatic, behavioral, and event-driven knowledge assets, five models are systematically compared: ARIMA, SARIMA, XGBoost, LSTM, and the proposed hybrid. The hybrid model dramatically outperforms all alternatives (RMSE = 10.68; MAE = 9.52 vs. RMSE = 89.29 for the next-best approach), demonstrating that architecturally informed knowledge integration produces forecasting gains that no standalone model can achieve. Beyond algorithmic performance, the study advances several theoretical contributions. First, it operationalizes Nonaka and Takeuchi's (1995) combination process, demonstrating that dividing cognitive labor between an explicit linear knowledge layer (ARIMA) and a quasi-tacit nonlinear correction layer (LSTM) constitutes a genuine knowledge recombination strategy. Second, it theorizes data preprocessing as a knowledge governance decision, arguing that the choice of what historical knowledge to retain, discount, or reconstruct is as strategically consequential for organizational learning as the choice of the algorithm to deploy. Third, it reframes multivariate dataset construction as the deliberate assembly of a structured organizational knowledge portfolio distinguishing explicit, behavioral, and tacit-made-explicit knowledge assets. Fourth, this research introduces post-disruption selective reconstruction as a generalizable organizational learning principle applicable beyond tourism to any knowledge-intensive sector navigating structural discontinuities. Implications are drawn for tourism authorities, IS practitioners, and knowledge managers seeking to institutionalize AI-driven foresight within learning organization frameworks.