Point forecast reconciliation of collection of time series with linear aggregation constraints has evolved substantially over the last decade. A few commonly used methods are GLS (generalized least squares), OLS (ordinary least squares), WLS (weighted least squares), and MinT (minimum trace). GLS and MinT have similar mathematical expressions, but they differ by the covariance matrix used. OLS and WLS can be considered as special cases of MinT where they differ by the assumptions made about the structure of the covariance matrix. All these methods ensure that the reconciled forecasts are unbiased, provided that the base forecasts are unbiased. The ERM (empirical risk minimizer) approach was proposed to relax the assumption of unbiasedness. This paper proves that (a) GLS and MinT reduce to the same solution; (b) on average, a method similar to ERM (which we refer to as MinT-U) can produce better forecasts than MinT (lowest total mean squared error) which is then followed by OLS and then by base; and (c) the mean squared error of each series in the structure for MinT-U is smaller than that for MinT which is then followed by that for either OLS or base forecasts. We show these theoretical results using a set of simulation studies. We also evaluate them using the Australian domestic tourism data set.
翻译:过去十年来,对时间序列收集时间序列和线性汇总限制的预测对时间序列的调和发生了很大变化。一些常用的方法有:GLS(一般最低平方)、OLS(一般最低平方)、WLS(加权最低平方)和MinT(最低微微微微微微微微),GLS和MinT具有类似的数学表达方式,但因使用的共变矩阵不同而有所不同。OLS和WLS可被视为MINT的特殊案例,它们与对共变矩阵结构的假设不同。所有这些方法都确保调和的预测是公正的,条件是基本预测是公正的。机构风险管理(尽可能最小风险最小的平方)方法是为了放松对公正性的假设。本文证明(a)GLS和MinT(最小最小最小微微微微微微微微微微微微微微微微微微微微微微微微微微微微微微微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小微小