Payment Volume Forecasting using Hierarchical Regression with SARIMA Errors : Payment Volume Forecasting using Hierarchical Regression with SARIMA Errors

Abstract: When forecasting financial transaction volumes in different markets, different markets often exhibit similar seasonality patterns and public holiday behavior. In this thesis, an attempt is made at utilizing these similarities to improve forecasting accuracy as compared to forecasting each market individually. Bayesian hierarchical regression models with time series errors are used on daily transaction data. When fitting three years of historic data for all markets, no consistent significant improvements in forecasting accuracy was found over a non-hierarchical regression model. When the amount of historic data was limited to less than one year for a single market, with the other markets having three years of historic data, the hierarchical model significantly outperformed both non-hierarchical and naive reference models on the market with limited historic data.