pith. sign in

arxiv: 2605.23858 · v1 · pith:ESB4H5MJnew · submitted 2026-05-22 · 📊 stat.AP

Anticipating Continued Global Fertility Decline via Neural Forecasting

Daniel Ciganda , Facundo Morini , Francisco Piriz , Henrik-Alexander Schubert , Ugofilippo Basellini , Mikko Myrskyl\"a This is my paper

Pith reviewed 2026-05-25 02:22 UTC · model grok-4.3

classification 📊 stat.AP
keywords fertility forecastingneural networksdemographic projectionslow fertilityrecurrent neural networkglobal trendsprediction intervalsBayesTFR comparison
0
0 comments X

The pith

A neural network using only historical fertility data projects wider exposure to low fertility by 2040 than the UN BayesTFR model.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces NeuralTFR, a recurrent neural network trained on harmonized fertility series from 196 countries and territories to learn cross-country patterns of demographic momentum. It produces point forecasts and multi-quantile prediction intervals without external covariates. On held-out data from 2009 to 2023 the model records lower errors than both a naive drift baseline and the UN Bayesian hierarchical model while keeping competitive uncertainty calibration. Forward projections to 2040 indicate more countries will reach low or very low fertility levels than BayesTFR anticipates, though the declines remain milder than those in the Global Burden of Disease study.

Core claim

NeuralTFR, an endogenous global forecasting framework based on a recurrent neural network, achieves lower point-forecast errors than BayesTFR on 2009-2023 data and points to broader exposure to low and very low fertility by 2040, 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.

What carries the argument

NeuralTFR, a recurrent neural network that pools cross-country historical fertility series to learn demographic momentum and generate empirical prediction intervals via multi-quantile regression.

If this is right

  • NeuralTFR records lower point-forecast errors than Naive Drift and BayesTFR on the 2009-2023 evaluation period.
  • Uncertainty calibration stays competitive with the UN benchmark model.
  • Projections to 2040 show more countries entering low and very low fertility than BayesTFR indicates.
  • The projected paths remain less extreme than Global Burden of Disease estimates.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the historical-data-only assumption holds, population planning may need to prepare for longer periods of low fertility rather than assuming quick recovery.
  • Adding post-2023 observations would provide a direct test of whether the learned momentum patterns continue.
  • The same recurrent-network pooling approach could be applied to other series such as mortality or migration where cross-country regularities exist.

Load-bearing premise

Historical fertility series alone contain sufficient learnable structure to generate reliable out-of-sample forecasts without external covariates or policy variables.

What would settle it

Actual fertility rates recorded after 2023 that track BayesTFR stabilization paths more closely than NeuralTFR's low-fertility projections would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.23858 by Daniel Ciganda, Facundo Morini, Francisco Piriz, Henrik-Alexander Schubert, Mikko Myrskyl\"a, Ugofilippo Basellini.

Figure 1
Figure 1. Figure 1: Schematic representation of the Sliding Window approach. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of point forecast evaluation metrics across the 192 test countries for the [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of probabilistic calibration and sharpness metrics across the 192 test [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative held-out trajectories where NeuralTFR better captures persistent or [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Held-out trajectories for the United States and selected Nordic countries, 2009–2023. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative held-out trajectories where NeuralTFR struggles under unprecedented [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Aggregate projection comparison for the 151-country comparison sample, 2025–2040. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Regional fertility thresholds in Europe and Asia at the projection endpoint. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Regional fertility thresholds in the Middle East, Oceania, the Americas, and Africa [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
read the original 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. 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. Evaluated on a held-out period (2009--2023), NeuralTFR achieves lower point-forecast errors than a Naive Drift baseline and BayesTFR, the United Nations' Bayesian Hierarchical Model, while maintaining competitive uncertainty calibration. In forward projections to 2040, NeuralTFR points to 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces NeuralTFR, a recurrent neural network model for global total fertility rate (TFR) forecasting. Trained on harmonized historical TFR series from 196 countries and territories, the model pools cross-country information to capture demographic momentum and generates empirical prediction intervals through multi-quantile regression. On a 2009-2023 hold-out period, NeuralTFR reports lower point-forecast errors than a Naive Drift baseline and the UN's BayesTFR model while maintaining competitive uncertainty calibration. Forward projections to 2040 indicate broader exposure to low and very low fertility than BayesTFR, though less severe than Global Burden of Disease estimates.

Significance. If the performance advantages hold under scrutiny, the work supplies a useful empirical benchmark in the fertility-projection literature by offering a purely endogenous, data-driven alternative to models that embed explicit recovery assumptions or external drivers. The multi-quantile approach for uncertainty is a constructive methodological choice. The central projection result—that NeuralTFR anticipates wider low-fertility exposure by 2040—would, if substantiated, weaken support for near-term stabilization relative to BayesTFR.

major comments (3)
  1. [Abstract] Abstract: the claim that NeuralTFR achieves lower point-forecast errors than BayesTFR on the 2009-2023 hold-out is presented without accompanying numerical error values, confidence intervals, or statistical tests; this omission prevents assessment of whether the reported advantage is practically meaningful or statistically reliable.
  2. [Methods] Methods (model description): the central projection claim—that NeuralTFR forecasts broader low-fertility exposure by 2040—rests on the assumption that historical TFR trajectories alone contain sufficient stationary structure for reliable extrapolation to 2040; no sensitivity analysis to potential regime shifts or external drivers is reported, leaving the long-horizon validity untested.
  3. [Evaluation] Evaluation section: the manuscript provides no details on network architecture, training procedure, quantile loss specification, or data harmonization steps; without these, the soundness of the reported out-of-sample performance cannot be verified and the comparison to BayesTFR remains non-reproducible.
minor comments (2)
  1. [Abstract] Abstract: the precise error metric (MAE, RMSE, or other) underlying the 'lower point-forecast errors' claim is not stated.
  2. The paper would benefit from an explicit statement of the number of countries/territories with complete series and any imputation rules applied.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for these constructive comments, which highlight opportunities to strengthen the clarity, reproducibility, and robustness of the manuscript. We respond to each major comment below and will incorporate revisions accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that NeuralTFR achieves lower point-forecast errors than BayesTFR on the 2009-2023 hold-out is presented without accompanying numerical error values, confidence intervals, or statistical tests; this omission prevents assessment of whether the reported advantage is practically meaningful or statistically reliable.

    Authors: We agree that the abstract would be strengthened by quantitative detail. In the revised manuscript we will insert the specific point-forecast error values (MAE or equivalent) achieved by NeuralTFR, the Naive Drift baseline, and BayesTFR on the 2009-2023 hold-out. If formal statistical tests were conducted we will report them; otherwise the comparison will be presented descriptively with the numerical results. revision: yes

  2. Referee: [Methods] Methods (model description): the central projection claim—that NeuralTFR forecasts broader low-fertility exposure by 2040—rests on the assumption that historical TFR trajectories alone contain sufficient stationary structure for reliable extrapolation to 2040; no sensitivity analysis to potential regime shifts or external drivers is reported, leaving the long-horizon validity untested.

    Authors: NeuralTFR is intentionally constructed as a purely endogenous benchmark that learns only from historical TFR series, deliberately avoiding recovery assumptions or external covariates that appear in other models. We recognize that this leaves long-horizon extrapolation sensitive to possible regime shifts. We will add a dedicated sensitivity subsection that retrains the model on truncated historical windows and examines forecast stability under simulated non-stationarities. revision: yes

  3. Referee: [Evaluation] Evaluation section: the manuscript provides no details on network architecture, training procedure, quantile loss specification, or data harmonization steps; without these, the soundness of the reported out-of-sample performance cannot be verified and the comparison to BayesTFR remains non-reproducible.

    Authors: We apologize for the lack of explicit detail. The Methods section already specifies an LSTM-based recurrent architecture, multi-quantile (pinball) loss, and harmonized UN/national TFR series, but these elements will be expanded with concrete hyperparameters (layers, hidden units, optimizer, epochs), the exact quantile loss formulation, and a step-by-step description of the data harmonization pipeline. Reproducible code or pseudocode will also be supplied. revision: yes

Circularity Check

0 steps flagged

NeuralTFR derives forecasts from trained RNN on historical data with no reduction by construction

full rationale

The paper trains a recurrent neural network on harmonized historical TFR series from 196 countries/territories, evaluates point-forecast errors and uncertainty calibration on a 2009-2023 hold-out period against baselines including BayesTFR, and then generates forward projections to 2040. No equations, model descriptions, or claims indicate that the 2040 projections or the claim of broader low-fertility exposure reduce by construction to quantities already fitted inside the model itself. The derivation chain is a standard supervised learning setup (endogenous RNN + multi-quantile regression) whose outputs are not definitionally equivalent to its training inputs. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling appear in the provided text. This is the most common honest non-finding for empirical forecasting papers.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that cross-country historical fertility trajectories encode generalizable momentum that a recurrent network can extract without external drivers; the network weights themselves are fitted parameters.

free parameters (1)
  • RNN weights, biases, and quantile parameters
    Learned during training on the panel data to produce the forecasts and intervals.
axioms (1)
  • domain assumption Historical fertility series from 196 countries contain learnable endogenous momentum sufficient for out-of-sample prediction
    Invoked by the description of the endogenous global forecasting framework.

pith-pipeline@v0.9.0 · 5740 in / 1217 out tokens · 38246 ms · 2026-05-25T02:22:46.654269+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages

  1. [1]

    Akiba, T., S. Sano, T. Yanase, T. Ohta, and M. Koyama (2019). Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pp.\ 2623--2631

    work page 2019

  2. [2]

    Alkema, L., A. E. Raftery, P. Gerland, S. J. Clark, F. Pelletier, T. Buettner, and G. K. Heilig (2011). Probabilistic projections of the total fertility rate for all countries. Demography\/ 48\/ (3), 815--839

    work page 2011

  3. [3]

    Bandara, K., P. Shi, C. Bergmeir, H. Hewamalage, Q. Tran, and B. Seaman (2019). Sales demand forecast in e-commerce using a long short-term memory neural network methodology. In T. Gedeon, K. W. Wong, and M. Lee (Eds.), Neural Information Processing , Cham, pp.\ 462--474. Springer International Publishing

    work page 2019

  4. [4]

    Sobotka, and K

    Basten, S., T. Sobotka, and K. Zeman (2014). Future fertility in low fertility countries. In W. Lutz, W. P. Butz, and S. KC (Eds.), World Population and Human Capital in the Twenty-First Century , pp.\ 39--146. Oxford University Press

    work page 2014

  5. [5]

    Beyaztas, U. and H. Shang (2022). Machine-learning-based functional time series forecasting: Application to age-specific mortality rates. Forecasting\/ 4\/ (1), 394--408

    work page 2022

  6. [6]

    u sener, and M. Myrskyl \

    Campisi, N., H. Kulu, J. Mikolai, S. Kl \"u sener, and M. Myrskyl \"a (2023). A spatial perspective on the unexpected nordic fertility decline: The relevance of economic and social contexts. Applied Spatial Analysis and Policy\/ 16\/ (1), 1--31

    work page 2023

  7. [7]

    Van Merri \"e nboer, C

    Cho, K., B. Van Merri \"e nboer, C . Gul c ehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio (2014). Learning phrase representations using rnn encoder--decoder for statistical machine translation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) , pp.\ 1724--1734

    work page 2014

  8. [8]

    Fischer, T. and C. Krauss (2018). Deep learning with long short-term memory networks for financial market predictions. European Journal of Operational Research\/ 270\/ (2), 654--669

    work page 2018

  9. [9]

    Global fertility in 204 countries and territories, 1950--2021, with forecasts to 2100: a comprehensive demographic analysis for the global burden of disease study 2021

    GBD 2021 Fertility and Forecasting Collaborators (2024). Global fertility in 204 countries and territories, 1950--2021, with forecasts to 2100: a comprehensive demographic analysis for the global burden of disease study 2021. The Lancet\/ 403\/ (10440), 2057--2099

    work page 2021

  10. [10]

    Gietel-Basten, S. and T. Sobotka (2020). Uncertain population futures: Critical reflections on the ihme scenarios of future fertility, mortality, migration and population trends from 2017 to 2100. OSF Preprints\/ 10

    work page 2020

  11. [11]

    Sobotka, et al

    Gietel-Basten, S., T. Sobotka, et al. (2020). Letter on ' Fertility , mortality, migration, and population scenarios for 195 countries and territories from 2017 to 2100: a forecasting analysis for the Global Burden of Disease Study ' by S. E. Vollset et al. https://osf.io/ytf6m/overview?view_only=5389a004f7e94917b30b5a4d4ea7a154

    work page 2020

  12. [12]

    Gneiting, T. and M. Katzfuss (2014). Probabilistic forecasting. Annual Review of Statistics and Its Application\/ 1\/ (1), 125--151

    work page 2014

  13. [13]

    Gneiting, T. and A. E. Raftery (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association\/ 102\/ (477), 359--378

    work page 2007

  14. [14]

    Goujon, A. (2025). The demographic future that we do not know about. Science\/ 390\/ (6725)

    work page 2025

  15. [15]

    Nis \'e n, V

    Hellstrand, J., J. Nis \'e n, V. Miranda, P. Fallesen, L. Dommermuth, and M. Myrskyl \"a (2021). Not just later, but fewer: Novel trends in cohort fertility in the nordic countries. Demography\/ 58\/ (4), 1373--1399

    work page 2021

  16. [16]

    Hochreiter, S. and J. Schmidhuber (1997). Long short-term memory. Neural computation\/ 9\/ (8), 1735--1780

    work page 1997

  17. [17]

    Global fertility, mortality, migration, and population forecasts 2017--2100

    Institute for Health Metrics and Evaluation (IHME) (2020). Global fertility, mortality, migration, and population forecasts 2017--2100

    work page 2020

  18. [18]

    Gasthaus, Y

    Januschowski, T., J. Gasthaus, Y. Wang, D. Salinas, V. Flunkert, M. Bohlke-Schneider, and L. Callot (2020). Criteria for classifying forecasting methods. international Journal of forecasting\/ 36\/ (1), 167--177

    work page 2020

  19. [19]

    Dhakad, M

    KC, S., M. Dhakad, M. Potan c okov \'a , S. Adhikari, D. Yildiz, M. Mamolo, T. Sobotka, K. Zeman, G. Abel, W. Lutz, et al. (2024). Updating the shared socioeconomic pathways (ssps) global population and human capital projections

    work page 2024

  20. [20]

    Kearney, M. S., P. B. Levine, and L. Pardue (2022). The puzzle of falling US birth rates since the Great Recession . Journal of Economic Perspectives\/ 36\/ (1), 151--176

    work page 2022

  21. [21]

    Koenker, R. and G. Bassett Jr (1978). Regression quantiles. Econometrica: journal of the Econometric Society\/ , 33--50

    work page 1978

  22. [22]

    Chang, Y

    Lai, G., W.-C. Chang, Y. Yang, and H. Liu (2018). Modeling long- and short-term temporal patterns with deep neural networks. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval , pp.\ 95--104

    work page 2018

  23. [23]

    Lutz, W., W. P. Butz, et al. (2014). World population and human capital in the twenty-first century . OUP Oxford

    work page 2014

  24. [24]

    Goujon, S

    Lutz, W., A. Goujon, S. Kc, M. Stonawski, and N. Stilianakis (2018). Demographic and human capital scenarios for the 21st century: 2018 assessment for 201 countries . Publications Office of the European Union

    work page 2018

  25. [25]

    Skirbekk, and M

    Lutz, W., V. Skirbekk, and M. R. Testa (2006). The low-fertility trap hypothesis: Forces that may lead to further postponement and fewer births in Europe . Vienna yearbook of population research\/ , 167--192

    work page 2006

  26. [26]

    Montero-Manso, P. and R. J. Hyndman (2021). Principles and algorithms for forecasting groups of time series: Locality and globality. International Journal of Forecasting\/ 37\/ (4), 1632--1653

    work page 2021

  27. [27]

    Kohler, and F

    Myrskyl \"a , M., H.-P. Kohler, and F. C. Billari (2009). Advances in development reverse fertility declines . Nature\/ 460\/ (7256), 741--743

    work page 2009

  28. [28]

    Levantesi, M

    Nigri, A., S. Levantesi, M. Marino, S. Scognamiglio, and F. Perla (2019). A deep learning integrated lee-carter model. Risks\/ 7\/ (1), 33

    work page 2019

  29. [29]

    Paranhos, L. (2025). Predicting inflation with recurrent neural networks. International Journal of Forecasting\/ . Article in press

    work page 2025

  30. [30]

    Park, P. (2025). Next-generation mortality forecasting with deep learning. OSF Preprints\/

    work page 2025

  31. [31]

    Raftery, A. E., L. Alkema, and P. Gerland (2014). Bayesian population projections for the united nations. Statistical science: a review journal of the Institute of Mathematical Statistics\/ 29\/ (1), 58

    work page 2014

  32. [32]

    Raftery, A. E. and H. S ev c \' kov \'a (2023). Probabilistic population forecasting: Short to very long-term. International Journal of Forecasting\/ 39\/ (1), 73--97

    work page 2023

  33. [33]

    Flunkert, J

    Salinas, D., V. Flunkert, J. Gasthaus, and T. Januschowski (2020). Deepar: Probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting\/ 36\/ (3), 1181--1191

    work page 2020

  34. [34]

    Smyl, S. (2020). A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting. International Journal of Forecasting\/ 36\/ (1), 75--85

    work page 2020

  35. [35]

    Susmann, H. and L. Alkema (2025). Flexible modelling of demographic transition processes with a bayesian hierarchical b-splines model. Journal of the Royal Statistical Society Series C: Applied Statistics\/ 74 , 1340--1371

    work page 2025

  36. [36]

    Vinyals, and Q

    Sutskever, I., O. Vinyals, and Q. V. Le (2014). Sequence to sequence learning with neural networks. Advances in neural information processing systems\/ 27

    work page 2014

  37. [37]

    World Population Prospects 2024: Summary of Results

    United Nations (2024). World Population Prospects 2024: Summary of Results . New York: United Nations Department of Economic and Social Affairs, Population Division

    work page 2024

  38. [38]

    Zheng, H., H. Wang, R. Zhu, and J.-H. Xue (2025). A brief review of deep learning methods in mortality forecasting. Annals of Actuarial Science\/ , 1--16

    work page 2025

  39. [39]

    International Journal of Forecasting , year =

    Paranhos, Livia , title =. International Journal of Forecasting , year =

    work page

  40. [40]

    Neural Information Processing , editor =

    Bandara, Kasun and Shi, Peng and Bergmeir, Christoph and Hewamalage, Hansika and Tran, Quang and Seaman, Bill , title =. Neural Information Processing , editor =. 2019 , address =

    work page 2019

  41. [41]

    European Journal of Operational Research , volume =

    Fischer, Thomas and Krauss, Christopher , title =. European Journal of Operational Research , volume =. 2018 , doi =

    work page 2018

  42. [42]

    International Journal of Forecasting , volume =

    Salinas, David and Flunkert, Valentin and Gasthaus, Jan and Januschowski, Tim , title =. International Journal of Forecasting , volume =. 2020 , doi =

    work page 2020

  43. [43]

    International Journal of Forecasting , volume =

    Smyl, Slawek , title =. International Journal of Forecasting , volume =. 2020 , doi =

    work page 2020

  44. [44]

    The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval , pages =

    Lai, Guokun and Chang, Wei-Cheng and Yang, Yiming and Liu, Hanxiao , title =. The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval , pages =. 2018 , doi =

    work page 2018

  45. [45]

    Science , volume =

    Goujon, Anne , title =. Science , volume =. 2025 , doi =

    work page 2025

  46. [46]

    and Levine, Phillip B

    Kearney, Melissa S. and Levine, Phillip B. and Pardue, Luke , title =. Journal of Economic Perspectives , volume =. 2022 , doi =

    work page 2022

  47. [47]

    A Spatial Perspective on the Unexpected Nordic Fertility Decline: The Relevance of Economic and Social Contexts , journal =

    Campisi, Natalie and Kulu, Hill and Mikolai, Julia and Kl. A Spatial Perspective on the Unexpected Nordic Fertility Decline: The Relevance of Economic and Social Contexts , journal =. 2023 , doi =

    work page 2023

  48. [48]

    Annual Review of Statistics and Its Application , volume =

    Probabilistic Forecasting , author =. Annual Review of Statistics and Its Application , volume =. 2014 , publisher =

    work page 2014

  49. [49]

    Journal of the American Statistical Association , volume =

    Strictly Proper Scoring Rules, Prediction, and Estimation , author =. Journal of the American Statistical Association , volume =. 2007 , publisher =

    work page 2007

  50. [50]

    Journal of the Royal Statistical Society Series C: Applied Statistics , volume =

    Susmann, Herbert and Alkema, Leontine , title =. Journal of the Royal Statistical Society Series C: Applied Statistics , volume =. 2025 , doi =

    work page 2025