Forecast accuracy hardly improves with method complexity when completing cohort fertility

Abstract

SignificanceInformation on cohort fertility is critical for the understanding of population dynamics, but only in historical settings can it be calculated without forecasting. Several forecasting methods exist, but their strengths and weaknesses have not been evaluated. Relying on the Human Fertility Database, the largest high-quality fertility dataset to date, and the globally representative United Nations World Population Prospects, we present an assessment of all major methods that complete cohort fertility. This analysis is crucial to advance the understanding of benefits and drawbacks of state-of-the-art methods. We analyze forecast accuracy and uncertainty quantification, identify methodological breakthroughs, and uncover unresolved issues. This study constitutes an evaluation benchmark for cohort fertility forecasting and may inspire establishment of similar evaluation benchmarks in related fields.

Preparing for the future: The changing demographic composition of hospital patients in Denmark between 2013 and 2050

“…Keeping baseline levels constant during the projection period is a pragmatic approach in forecasting and has been applied to detailed projections of hospital care use [28] or fertility [31]. Especially when over-arching trends are difficult to predict, freezing rates has been shown to outperform statistically sophisticated techniques [31]. Additionally, the intention of this study is not to accurately forecast hospital care use in future for Denmark, but to provide a glimpse into the relative contribution of population aging to the total hospital care use by 2050, given that hospitalization patterns remain constant over the forecasting period.…”

Synergy in fertility forecasting: improving forecast accuracy through model averaging

“…Though the level of fertility in industrialised countries has at times been of major concern (e.g., Kuczynski (1937); Booth (1986); Longman (2004)), models for forecasting fertility have been developed relatively recently (for an earlier review, see Booth (2006)). Indeed, many of the models used in fertility forecasting were originally designed to smooth age-specific fertility rates or complete the incomplete experience of a single cohort, and relatively few have been used in longerterm forecasting of period fertility rates; further, many are deterministic providing no indication of uncertainty (Bohk-Ewald et al 2018).…”

“…The strengths and weaknesses of fertility forecasting models have not been thoroughly evaluated. Noting the absence of guidance on model choice, Bohk-Ewald et al (2018) compared the point and interval forecast accuracy of 20 major models for fertility forecasting with 162 variants. In the context of completing cohort fertility, their evaluation found that only four methods were consistently more accurate than the constant (no change) model and, among these, complex Bayesian models did not outperform simple extrapolative models.…”

The maximum entropy mortality model: forecasting mortality using statistical moments

“…We are also interested in showing that accurate forecasts of age-specific mortality levels can be obtained using statistical moments and the information provided by the age-at-death distribution; • compare the MEM against other well established mortality models and determine its newly added value; • validate the MEM predictions against a benchmark, that is, a simplistic trend extrapolation of the age-specific death rates (naïve model) in order to justify the increase in complexity of the proposed method. This objective is justified and inspired by Bohk-Ewald et al (2018), a study where 20 major fertility forecasting methods are evaluated. The main findings show that across multiple measures of fertility forecast accuracy only four methods consistently outperform the naïve model.…”