Bayesian Forecasting for Seemingly Unrelated Time Series: Application to Local Government Revenue Forecasting
提出一种结合多状态卡尔曼滤波和条件独立分层方法的贝叶斯预测方法C-MSKF，用于处理看似无关但受共同外部因素影响的多组时间序列数据。以宾夕法尼亚州阿勒格尼县40个学区的所得税收入预测为例，证明C-MSKF比单变量MSKF更准确，且历史数据越短、预测期越长、学区对经济周期越敏感时优势越明显。
One important implementation of Bayesian forecasting is the Multi-State Kalman Filter (MSKF) method. It is particularly suited for short and irregular time series data. In certain applications, time series data are available on numerous parallel observational units which, while not having cause-and-effect relationships between them, are subject to the same external forces (e.g., business cycles). Treating them separately may lose useful information for forecasting. For such situations, involving seemingly unrelated time series, this article develops a Bayesian forecasting method called C-MSKF that combines the MSKF method with the Conditionally Independent Hierarchical method. A case study on forecasting income tax revenue for each of forty school districts in Allegheny County, Pennsylvania, based on fifteen years of data, is used to illustrate the application of C-MSKF in comparison with univariate MSKF. Results show that C-MSKF is more accurate than MSKF. The relative accuracy of C-MSKF increases with decreasing length of historical time series data, increasing forecasting horizon, and sensitivity of school districts to the economic cycle.