All Times ET
Keywords: Time Series Forecasting, Hierarchical Time Series, ARIMA, LSTM, Odds, Stacked LSTM, BiDirectional LSTM, CNN LSTM, Convolutional LSTM
This paper addresses a common problem with hierarchical time series. Time series analysis demands the series for a model to be the sum of multiple series at corresponding sub-levels. Hierarchical Time Series presents a two-fold problem. First, each individual time series model at each level in the hierarchy must be estimated separately. Second, those models must maintain their hierarchical structure over the specified period of time, which is complicated by performance degradation of the higher-level models in the hierarchy. This performance loss is attributable to the summation of the bottom-level time series models. In this paper, the proposed methodology works to correct this degradation of performance through a top-down approach using odds, time series and systems of linear equations. Vertically, the total counts of corresponding series at each sub-level are captured while horizontally odds are computed to establish and preserve the relationship between each respective time series model at each level. The results, based on root mean square percentage error with simulated hierarchical time series data, are promising.