Title: A hybrid machine learning framework for tax revenues monitoring Authors:  Daria Scacciatelli - Sogei (Italy) [presenting]
Eugenio Cangiano - Sogei (Italy)
Francesco De Napoli - Sogei (Italy)
Abstract: Monitoring tax revenues at least monthly is extremely important for assessing the convergence of public finance figures with annual objectives. This is especially relevant due to the potential revisions in annual budget forecasts by the Italian Ministry of Economy and Finance, influenced by changes in fiscal policy measures or updates in macroeconomic scenarios. To assess the impact of such revisions, a higher-frequency model is proposed, incorporating additional information gathered throughout the year. The proposed hybrid machine learning framework, named HGB, rooted in the gradient boosting algorithm, is designed to generate short-term forecasts of tax revenues. This framework integrates feature selection methods, auto-regressive models, and Machine Learning regression algorithms. Data from diverse sources are gathered and directed to a centralized data hub, where the Boruta algorithm identifies relevant information. The SARIMA model predicts future values for selected variables, and the XGBoost model uses these predictions to derive tax revenue forecasts. In the experimental results, the analysis focuses on excise duty on mineral oil, representing one of the indirect taxes. The HGB framework exhibited high predictive accuracy, outperforming traditional autoregressive models used as benchmarks. The evaluation was performed using the k-fold cross-validation method.