Forecast combination based forecast reconciliation: insights and extensions

TL;DR

This paper extends the forecast combination-based forecast reconciliation approach, introducing new mathematical insights, considering endogenous constraints, and evaluating the methods on tourism demand data to improve forecast accuracy.

Contribution

It provides a detailed mathematical derivation of the LCC reconciliation formula, explores endogenous constraints, and assesses the methods' performance on real-world data.

Findings

LCC reconciliation is derived as a quadratic loss minimization.

Endogenous constraints allow coherent revisions of all series.

Forecast pooling improves accuracy in the Australian Tourism dataset.

Abstract

In a recent paper, while elucidating the links between forecast combination and cross-sectional forecast reconciliation, Hollyman et al. (2021) have proposed a forecast combination-based approach to the reconciliation of a simple hierarchy. A new Level Conditional Coherent (LCC) point forecast reconciliation procedure was developed, and it was shown that the simple average of a set of LCC, and bottom-up reconciled forecasts (called Combined Conditional Coherent, CCC) results in good performance as compared to those obtained through the state-of-the-art cross-sectional reconciliation procedures. In this paper, we build upon and extend this proposal along some new directions. (1) We shed light on the nature and the mathematical derivation of the LCC reconciliation formula, showing that it is the result of an exogenously linearly constrained minimization of a quadratic loss function in the differences between the target and the base forecasts with a diagonal associated matrix. (2) Endogenous constraints may be considered as well, resulting in level conditional reconciled forecasts of all the involved series, where both the upper and the bottom time series are coherently revised. (3) As the LCC procedure does not guarantee the non-negativity of the reconciled forecasts, we argue that - when non-negativity is a natural attribute of the variables to be forecast - its interpretation as an unbiased top-down reconciliation procedure leaves room for some doubts. (4) The new procedures are used in a forecasting experiment on the classical Australian Tourism Demand (Visitor Nights) dataset. Due to the crucial role played by the (possibly different) models used to compute the base forecasts, we re-interpret the CCC reconciliation of Hollyman et al. (2021) as a forecast pooling approach, showing that accuracy improvement may be gained by adopting a simple forecast averaging strategy.