Since the conjecture of Richards (2008), the work by Cairns et al. (2016) and subsequent developments by Boumezoued (2020), Boumezoued et al. (2020) and Boumezoued et al. (2019), it has been acknowledged that observations from censuses have led to major problems of reliability in estimates of general population mortality rates as implemented in practice. These issues led to mis-interpretation of some key mortality characteristics in the past decades, including ”false cohort effects”. To overcome these issues, the exposure estimates for a given country can be corrected by using monthly fertility records. However, in the absence of birth-by-month data, the recent developments are not applicable. Therefore, this paper explores new solutions regarding the construction of mortality tables in this context, based on machine learning techniques. As a main result, it is demonstrated that the new exposure models proposed in this paper allow to provide correction with high quality and to improve the fitting of stochastic mortality models without a cohort component, as it is the case for the existing correction method based on monthly fertility data.

自理查兹（2008年）的猜想以来，凯恩斯（Cairns）等人的工作。（2016）和Boumezoued（2020）的后续发展，Boumezoued等人。（2020年）和Boumezoued等人。（2019年），人们已经认识到，从人口普查中观察到的结果导致在实践中对总体人口死亡率进行估算时存在重大可靠性问题。这些问题导致过去几十年来对某些关键死亡率特征的误解，包括“虚假队列效应”。为了克服这些问题，可以通过使用每月生育率记录来更正给定国家的暴露估计数。但是，在没有按月出生数据的情况下，最近的事态发展是不适用的。因此，本文基于机器学习技术，探索了在此背景下构建死亡率表的新解决方案。