Anticipating Continued Global Fertility Decline via Neural Forecasting

arXiv:2605.23858 · stat.AP · Submitted 2026-05-22 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Anticipating Continued Global Fertility Decline via Neural Forecasting".

Jane: Here's the gist of the paper. The paper introduces NeuralTFR, an endogenous global forecasting framework based on a recurrent neural network to empirically benchmark whether countries currently experiencing rapid fertility decline are approaching stabilization or entering a more persistent low-fertility regime This framework pools cross-country information to learn demographic momentum and generate empirical prediction intervals via multi-quantile regression.

Tom: The Problem Addressed The accelerating shift toward low and ultra-low fertility has intensified the debate over whether countries now undergoing rapid decline are approaching stabilization or entering a more persistent low-fertility regime.

Jane: Existing projection systems answer that question differently because they embed different assumptions about recovery and about the role of external drivers.

Tom: The authors introduce NeuralTFR to provide an empirical benchmark in this debate.

Paper summary: Tom: So, wrapping up this discussion about the paper "Anticipating Continued Global Fertility Decline via Neural Forecasting," what's the main implication for anyone listening?

Jane: It suggests that relying on a single projection system, like just the United Nations model or BayesTFR, might be missing some of the complex dynamics of how fertility rates actually behave globally over time.

Lu: NeuralTFR provides an empirical benchmark that shows where those existing systems fall short when it comes to predicting whether we're seeing a stabilization or a persistent low-fertility zone.

Meng: From a deployment standpoint, the idea is valuable for understanding the uncertainty, but since it has to learn from historical data, running this at scale on real traffic will be tricky because of those historical constraints.

Tom: The title itself really sets up the tension there, doesn't it? "Anticipating Continued Global Fertility Decline via Neural Forecasting." It’s about using neural networks to predict the future path of fertility.

Jane: Exactly. It moves the discussion away from just picking one institutional approach and gives us a way to see how different models handle those complex transitions in birth rates.

Lu: The authors are showing that pooling cross-country data through this recurrent neural network structure lets the system learn non-linear dynamics that weren't explicitly programmed into older, more phase-based models.

Meng: It’s a useful tool for seeing what happens when you try to be truly endogenous in your forecasting, but you have to be careful about those historical limits they mentioned.

Tom: So the paper gives us a new way to test these big questions about global population trends using machine learning on demographic data.

Conclusion: Tom: So, this paper is about NeuralTFR, which they call an endogenous global forecasting framework using a recurrent neural network to check if countries with fast fertility decline are stabilizing or just going into a long low-fertility zone.

Jane: It's trying to answer that big debate about whether fertility declines are temporary pauses or the start of something much more permanent and persistent.

Lu: The authors pool data from one hundred ninety-six countries using this neural network to learn how demographic momentum works across different series, which is a big move because it learns those shared dynamics instead of just applying separate rules for each country.

Meng: From a deployment standpoint, the idea of pooling information sounds cool, but I'm wondering how much real-world noise is in that historical data they used to train the model.

Lalam: The model uses a gated recurrent unit architecture to handle those time dependencies, and they use non-parametric quantile regression to get prediction intervals for uncertainty.

Tom: The evaluation on the held-out period from two thousand nine to two thousand twenty-three shows that NeuralTFR actually has lower point forecast errors than a naive drift baseline or the BayesTFR model.

Jane: So, what does that mean for us? It suggests this neural approach gives us a broader look at low and very low fertility than some of the existing United Nations projections do.

Lu: They found that when fertility decline continues or picks up after a temporary pause, NeuralTFR captures those dynamics better than the BayesTFR model did.

Meng: But they also admitted a limitation: since it only learns from historical trajectories, it can't predict anything outside of what has already happened in the data set.

Tom: And they flagged that it struggles more with sudden reversals in places like Central Asia compared to countries like South Korea's continued decline toward ultra-low fertility.

Jane: So, the authors are basically saying this framework is better for medium-run forecasts than for really long-term projections because of how much history it needs to learn from.

Lalam: This paper shows that pooling cross-country information through this kind of structure can empirically benchmark whether we are seeing a stabilization or just a persistent low-fertility regime.

Tom: It gives us a new tool for testing these big questions about global population trends, and now we have to figure out if this neural approach is robust enough for real-time monitoring.

Jane: Right, so it's not just another projection model; it’s a way to see the underlying momentum that might be missed in simpler systems.

Max Planck Institute for Demographic Research · Statistics Institute, University of the Republic of Montevideo, Uruguay · Digital Sense, Montevideo, Uruguay · Helsinki Institute for Demography and Population Health, University of Helsinki · Max Planck – University of Helsinki Center for Social Inequalities in Population Health

stat.AP

Submitted: 2026-05-22

Updated: 2026-10-29

Comments: Revised following peer review. Expanded dataset, additional regression benchmarks, and new interpretability analyses

Code: https://github.com/FaMori/NeuralTFR

Project page: https://famori.github.io/NeuralTFR

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 80/100

The gist: The paper introduces NeuralTFR, an endogenous global forecasting framework based on a recurrent neural network to empirically benchmark whether countries currently experiencing rapid fertility…

Key concepts

Endogenous Global Forecasting Framework
This is a system built using a recurrent neural network that learns demographic trends directly from the observed data of fertility rates. Instead of relying on fixed assumptions, it generates its own predictions based on the historical patterns it has learned across many countries.
Cross-Series Learning Strategy
The model trains a single AI simultaneously on fertility data from 196 different countries. This allows the system to learn shared, non-linear dynamics and momentum in fertility transitions by comparing and harmonizing these diverse historical time series together.
Quantile Regression for Uncertainty
Instead of giving a single point prediction, the model uses non-parametric quantile regression to generate prediction intervals. This helps quantify the uncertainty around its forecasts, showing a range of likely outcomes rather than just one single number.

Terminology

Summary

The paper introduces NeuralTFR, an endogenous global forecasting framework based on a recurrent neural network to empirically benchmark whether countries currently experiencing rapid fertility decline are approaching stabilization or entering a more persistent low-fertility regime This framework pools cross-country information to learn demographic momentum and generate empirical prediction intervals via multi-quantile regression. p. 0

Where the paper says it

To provide an empirical benchmark in this debate, we introduce NeuralTFR, an endogenous global forecasting framework based on a recurrent neural network. Drawing on a harmonized panel of historical fertility series from 196 countries and territories, the model pools cross-country information to learn demographic momentum and generate empirical prediction intervals via multi-quantile regression.

The Problem Addressed

The accelerating shift toward low and ultra-low fertility has intensified the debate over whether countries now undergoing rapid decline are approaching stabilization or entering a more persistent low-fertility regime. Existing projection systems answer that question differently because they embed different assumptions about recovery and about the role of external drivers. p. 0

Where the paper says it

Anticipating Continued Global Fertility Decline via Neural Forecasting Daniel CigandaB1,2, Facundo Morini2 , Francisco Piriz3 , Henrik-Alexander Schubert1,5, Ugofilippo Basellini1 , and Mikko Myrskylä1,4,5 1Max Planck Institute for Demographic Research, Rostock, Germany 2Statistics Institute, University of the Republic, Montevideo, Uruguay 3Digital Sense, Montevideo, Uruguay 4Helsinki Institute for Demography and Population Health, University of Helsinki, Helsinki, Finland 5Max Planck – University of Helsinki Center for Social Inequalities in Population Health, Rostock, Germany; Helsinki, Finland Abstract The accelerating shift toward low and ultra-low fertility has intensified the debate over whether countries now undergoing rapid decline are approaching stabilization or entering a more persistent low-fertility regime. Existing projection systems answer that question differently because they embed different assumptions about recovery and about the role of external drivers.

The authors introduce NeuralTFR to provide an empirical benchmark in this debate. p. 0
Where the paper says it

Existing projection systems answer that question differently because they embed different assumptions about recovery and about the role of external drivers. To provide an empirical benchmark in this debate, we introduce NeuralTFR, an endogenous global forecasting framework based on a recurrent neural network.

How the training signal is built

The model formulates demographic projection as an autoregressive sequence-to-sequence task for forecasting multiple steps ahead. It uses a cross-series learning strategy by jointly training a single model on a harmonized panel of historical fertility series from 196 countries and territories. p. 1

Where the paper says it

Rather than fitting isolated time-series models for each country, it uses a cross-series learning strategy. By jointly training a single model on a harmonized panel of historical fertility series from 196 countries and territories, NeuralTFR extracts shared, non-linear dynamics of the fertility transition directly from observed trajectories.

The data preprocessing involves applying two transformations: first, a natural logarithmic transformation is applied to the raw TFR values to compress the data range Second, a global standardization strategy is implemented where a standard scaler is fitted on the entire training set to compute a global mean (µglobal) and standard deviation (σglobal). p. 4
Where the paper says it

2 Preprocessing and Global Standardization Prior to model input, we apply two transformations to the raw TFR series to stabilize the learning process and facilitate convergence. First, a natural logarithmic transformation is applied to the raw TFR values to compress the data range.

Temporal structuring uses a sliding window approach where the Encoder Sequence is a historical context window of length Lenc defined as Xi,t = ⌈zi,t−Lenc+1,. p. 4
Where the paper says it

3 Temporal Structuring via Sliding Windows Once transformed, the time series are converted into input-output pairs suitable for supervised learning using a sliding window approach. For a given geographic entity i and a specific time step t, the data is partitioned into two sequences: • Encoder Sequence (Input): A historical context window of length Lenc defined as Xi,t = {zi,t−Lenc+1, .

.., zi,t⌉.

Where the gains come from

The framework's strength lies in its ability to pool information across related series by learning shared, non-linear dynamics of the fertility transition directly from observed trajectories. The model utilizes a Gated Recurrent Unit (GRU) architecture to capture temporal dependencies while mitigating the vanishing gradient problem. p. 6

Where the paper says it

However, standard RNNs struggle to learn long-range dependencies due to the vanishing gradient problem (Hochreiter and Schmidhuber, 1997). To mitigate this, our architecture utilizes the Gated Recurrent Unit (GRU) introduced by Cho et al.

Uncertainty quantification is achieved through non-parametric quantile regression by minimizing the multiquantile loss Ltotal. p. 7
Where the paper says it

This design allows for the direct construction of prediction intervals at any chosen nominal level, provided the corresponding quantiles are estimated, without imposing strict parametric assumptions, such as normality, on the error distribution. To optimize the network for probabilistic calibration, we train the model by minimizing the multiquantile loss (Koenker and Bassett Jr, 1978).

Furthermore, an ensemble strategy is implemented where the final probabilistic forecast for any given country is generated by taking the element-wise median of 10 structurally identical models. p. 8
Where the paper says it

To stabilize the forecasts and account for the model’s sensitivity to initial parameters, we implemented an ensemble strategy. The final probabilistic forecast for any given country is generated by taking the element-wise median of the quantile predictions from 10 structurally identical models, each trained independently utilizing a different random weight initialization seed and slightly varied hyperparameters to ensure model diversity.

Comparative Performance and Findings

Evaluation on a held-out period (2009–2023) shows that NeuralTFR achieves lower point-forecast errors than a Naive Drift baseline and BayesTFR. Substantively, the results suggest broader exposure to low and very low fertility than the United Nations projects, while still falling short of the most severe decline paths predicted by the Global Burden of Disease project. p. 1

Where the paper says it

Second, substantively, we provide a purely data-driven assessment of global fertility trajectories out to 2040. Our results suggest broader exposure to low and very low fertility than the United Nations projects, while still falling short of the most severe decline paths projected by the GBD framework.

The model demonstrates that when Fertility decline persists or resumes after a temporary pause, NeuralTFR better captures these dynamics compared to BayesTFR. p. 9
Where the paper says it

As illustrated by the distribution of point metrics across the test set (Figure 2), our proposed NeuralTFR model demonstrates robust generalization capabilities. For RMSE, sMAPE, and RMSSE, the median error for NeuralTFR is visibly lower than that of 2Eritrea and the Syrian Arab Republic have no observed values after 2008; Kosovo and Jersey could not be obtained from the BayesTFR output files.

Limitations and Future Directions

The main limitation of a purely endogenous model is that it can only learn from the range of trajectories present in the historical record. The framework remained too cautious in the face of South Korea’s continued collapse toward ultra-low fertility and struggled more with abrupt reversals in parts of Central Asia. p. 0

Where the paper says it

Anticipating Continued Global Fertility Decline via Neural Forecasting Daniel CigandaB1,2, Facundo Morini2 , Francisco Piriz3 , Henrik-Alexander Schubert1,5, Ugofilippo Basellini1 , and Mikko Myrskylä1,4,5 1Max Planck Institute for Demographic Research, Rostock, Germany 2Statistics Institute, University of the Republic, Montevideo, Uruguay 3Digital Sense, Montevideo, Uruguay 4Helsinki Institute for Demography and Population Health, University of Helsinki, Helsinki, Finland 5Max Planck – University of Helsinki Center for Social Inequalities in Population Health, Rostock, Germany; Helsinki, Finland Abstract The accelerating shift toward low and ultra-low fertility has intensified the debate over whether countries now undergoing rapid decline are approaching stabilization or entering a more persistent low-fertility regime. Existing projection systems answer that question differently because they embed different assumptions about recovery and about the role of external drivers.

A related limitation concerns the forecast horizon because extending it substantially would reduce the number of usable sliding windows. p. 5
Where the paper says it

, zi,t+Lpred } must be observed in the historical data. Increasing Lpred would therefore reduce the number of valid forecast origins and, importantly, remove the most recent years from model training, precisely where much of the new low- and ultra-low-fertility experience is observed.

The present framework is better suited to relatively short- and medium-run horizons than to very long-term projections. p. 3
Where the paper says it

For instance, Nigri et al. (2019) integrated Long Short-Term Memory (LSTM) networks into the classic Lee-Carter framework to forecast the mortality time index, demonstrating that recurrent architectures capture non-linear trends much more effectively than standard ARIMA processes.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Endogenous Global Forecasting Framework NeuralTFR can generate empirical prediction intervals by pooling cross-country information to learn demographic momentum directly from observed trajectories, achieving lower point-forecast errors than a Naive Drift baseline and BayesTFR.

  2. The system can provide forward projections to 2040 that suggest "broader exposure to low and very low fertility than BayesTFR, suggesting weaker support for near-term stabilization while still falling short of the most severe decline paths predicted by the Global Burden of Disease project."

  3. NeuralTFR's architecture, based on a recurrent neural network with Gated Recurrent Units (GRUs) and a sequence-to-sequence structure, can map historical input sequences to future projections autoregressively.

  4. The model utilizes non-parametric quantile regression to generate probabilistic forecasts, allowing it to construct prediction intervals without relying on restrictive parametric assumptions or the computational burden of traditional Bayesian models.

  5. The system can provide a robust assessment of uncertainty by using a dual approach involving non-parametric quantile regression and ensembling, resulting in probabilistic forecasts that are well calibrated and maintain a useful balance between coverage and sharpness across countries.

  6. The system can be used to assess the dependence of long-run fertility projections on assumptions by serving as an empirical benchmark for assessing how strongly long-run fertility projections depend on the assumptions built into those models.

Abstract

As fertility falls across much of the world, major projection systems disagree about whether decline will give way to recovery, reflecting different assumptions about post-transitional fertility. We ask instead what the historical record suggests, using NeuralTFR, a recurrent neural network that learns fertility dynamics solely from pooled histories of 230 countries and territories. In out-of-sample forecasts from three origins, it predicts fifteen years ahead as accurately as the United Nations' Bayesian model, and its prediction intervals score slightly better. Its forecasts reveal a simple nonlinear relationship, which a spline regression reproduces with similar accuracy: fertility tends to rise slightly from the lowest levels and decline from higher ones, with recent trends affecting the size of these changes more than their direction. By 2038, these dynamics leave very low fertility, below 1.5 births per woman, more widespread than the United Nations expects, without the extreme lows of the most pessimistic projections.

Related papers