Forecasting airline revenue across markets with machine learning

Abstract: Sales forecasting is an important aspect of business strategy, as it helps companies make informed decisions for their business development. With the rise of machine learning and AI, companies are using statistical and machine learning models to forecast short-term and long-term sales more accurately. In addition, it is important for companies to be able to estimate the impact of advertising on sales for different markets in order to optimize ad campaigns and make informed decisions about their advertising strategies. To address the former problem, ARIMA and LightGBM models with and without feature selection using Lasso regression are used to forecast revenue of an airline across different markets. The performance of the models is measured using Mean Absolute Error (MAE) and Symmetric Mean Absolute Percentage Error (sMAPE) metrics and both of them are compared to a simpler naïve approach called baseline. As for the impact of advertising, CausalImpact package is used to estimate the effect of non-advertising on revenue for each market for a specific period. LightGBM models without feature selection performed better in terms of error metrics for the majority of markets with improvements ranging between 41%-77% for sMAPE score and 30%-81% for MAE score when compared with baseline approach. While feature selection aided some markets in achieving lower error scores, it did not consistently improve performance across all markets. As for the effect of non-advertising on revenue, the analysis could not show a negative effect on revenues in all markets. It is difficult to say whether the effect was solely positive or negative for every market, because the results were not statistically significant for most markets, except one.