Conference Agenda

This paper extends the Lee-Carter model to include vanishing jumps on mortality rates, where
the impact is highest in the beginning and then gradually diminishes. Existing models either account
for transitory jumps over one period or permanent shifts. However, there is no literature on estimat-
ing mortality time series with jumps that have a vanishing effect over a small number of periods,
as is typically observed in pandemics. Using COVID-19 data, our proposed Bayesian model out-
performs transitory shock models in terms of in-sample fit, thus providing a more comprehensive
representation of mortality rates during a pandemic.

The Nordic countries have experienced rapid and unexpected fertility declines in the last decade. Future paths of fertility are a key input when charting the sustainability of social security systems. Hence realistic views of possible future paths of fertility, including the uncertainty regarding these paths, is critically important for economic and social planning in the Nordic countries.
An almost universal finding of probabilistic forecasting is that model-based prediction intervals are too narrow. The result is a misleadingly confident view of the future. We approach the problem of prediction uncertainty by generating empirical prediction intervals for our forecasts of Nordic period fertility up to 2050.
We derive the empirical prediction intervals for our forecasts from an out-of-sample cross-validation, utilizing historic forecasting errors. These quantify how uncertain we should be about the forecasts given the demonstrated forecast performance of the past. We then validate the calibration of the empirical prediction intervals and compare them to the performance of model-based prediction intervals. We use the following forecast models applied to data from the Human Fertility Database: First, a na"{i}ve approach fixing the last observed Total Fertility Rate (TFR). Second, a GARCH model of the TFR. Third, the Lee-Carter method for age-specific fertility rates. And fourth, a probabilistic, random-walk based forecast around the assumption that the ongoing fertility postponement would gradually slow down and come to an end by 2050.
In addition to presenting forecasts of Nordic fertility up to 2050, we expect to find that the empirical prediction intervals provide a more realistic view of the forecast uncertainty, exhibiting wider prediction intervals in comparison to the model-based prediction intervals.

After preliminary attempts at the end of the 20th century to compute probabilistic population forecasts (by Cohen, Keyfitz, Lee & Tuljapurkar, Alho, and others), the field became fully developed in the past two decades. I give a brief overview of the various applications of probabilistic demographic forecasts, spanning from national, subnational, and multi-country populations to the labour market, households, immigrants, and long-term care. I sketch the development from a frequentist to a Bayesian approach. Finally, I evaluate ex-post facto the predictive quality of selected population and household forecasts for Norway, the Netherlands, and France.

Population projections for small geographical areas are challenging even when data availability is good. Despite the presence of register data in Norway the current municipality level population projections by Statistics Norway are not satisfactory and are in the process of being replaced from a cohort-component framework to microsimulation. We propose a simpler and generalizable approach for downscaling national level population projections into municipality level projections, leveraging Norwegian register data and other data sources using an innovative neural network-based machine learning model. An additional advantage of this downscaling approach is that additional dimensions can easily be added to sub-national projections. We show this by disaggregating the Wittgenstein Centre population projections. The machine learning model is also trained by categorizing municipalities by special economic activities that might affect the population structure in that area. Such activities are the presence of fish farming, oil production, universities, or a high concentration of agricultural production.