Forecast av efterfrågan på flygsökningar för marknadsföringseffektivitet : En jämförelse mellan statistiska- och maskininlärningsmodeller

Accurate forecasting of online flight-search demand enables travel platforms to optimize marketing expenditures, enhance decision-making, and drive organizational value. This study evaluates two types of statistical time series models (Seasonal Autoregressive Integrated Moving Average and Prophet) alongside state-of-the-art machine learning techniques (Long Short-Term Memory networks and Light Gradient Boosting Machine) using daily click data from Flygresor.se (April 2016– December 2024). Systematic grid-search hyperparameter tuning and engineered calendar, trend, and lag features underlie model development. Forecast accuracy was assessed via root mean squared error and mean absolute percentage error with robustness tested by inclusion and exclusion of COVID-19 anomalies. LightGBM achieved the lowest overall MAPE, but both LSTM and LightGBM showed similar performance and converged to the actual monthly outcome considerably faster than the statistical models. Integration readiness and business value were examined through a Technology Acceptance Model interview and cost–benefit analysis, revealing that implementation costs are outweighed by projected profit gains. Findings demonstrate that advanced nonlinear forecasting models improve predictive performance and, through more accurate forecasts, deliver organizational benefits such as greater marketing-spend efficiency and increased decision confidence.