Comparative Analysis of Time Series Foundation Models for Demographic Forecasting: Enhancing Predictive Accuracy in US Population Dynamics

Aditya Akella, Jonathan FaraharXiv 2025 · DOI · arXiv

Introduction

Accurate demographic forecasting is essential for urban planning, healthcare resource allocation, and economic policy. However, predicting population dynamics is challenging due to non-linear "shocks" from migration or policy changes. Traditional statistical models like ARIMA often struggle with these irregularities, while deep learning models like LSTMs typically require extensive domain-specific data to generalize effectively. This challenge is particularly acute for minority populations, where historical records are often sparse or irregular.

This study investigates whether Time Series Foundation Models (TSFMs)—pre-trained on vast, diverse datasets—can overcome the "sparse data" problem. Specifically, we evaluate TimesFM, a large-scale decoder-only transformer model, against traditional baselines in forecasting US demographic shifts.

Methodology

The research compares four distinct modeling approaches:

  1. TimesFM: A foundation model with 200 million parameters, pre-trained on 100 billion real and synthetic time points.
  2. LSTM: A two-layer Long Short-Term Memory network, representing standard deep learning sequence models.
  3. ARIMA: A classic statistical model relying on linear patterns and stationarity.
  4. Linear Regression: A simple trend-line extrapolation used as a baseline.

The dataset spans 1990 to 2022, utilizing Federal Reserve Economic Data (FRED) for 1990–2019 and U.S. Census Bureau estimates for 2020–2022. The study focuses on six states (Alabama, California, Hawaii, New York, Texas, and Wyoming) and five racial groups.

Actual vs Predicted White Alone Population in Alabama (1990-2022) showing the model's ability to track historical trends and validation data.

Results and Discussion

The comparative analysis demonstrates the significant advantages of the foundation model approach. TimesFM achieved the lowest Mean Squared Error (MSE) in the vast majority of test cases, proving particularly robust for minority demographics.

Performance in Sparse Data Scenarios

A key finding is TimesFM’s ability to "democratize" accuracy for groups with limited historical data. For instance, in the Native Hawaiian demographic in New York, TimesFM reduced the MSE by three orders of magnitude compared to the LSTM baseline. This suggests that the model's pre-training on diverse temporal patterns allows it to understand the underlying "grammar" of time series evolution, even when specific local data is sparse.

Robustness to Trend Shifts

Traditional models often fail to account for sudden demographic shifts. While ARIMA and Linear Regression tend to project existing trends linearly, TimesFM demonstrated a superior ability to adjust to new downward or upward trajectories. This was notably observed in California's population data, where TimesFM maintained high precision despite sharp declines in 2019.

Comparative MSE results across different states and racial groups, highlighting TimesFM's superior performance in most categories.

As shown in the table above, TimesFM consistently achieves lower error rates across various states and racial categories compared to traditional deep learning and statistical methods.

Conclusion

TimesFM represents a significant advancement in demographic forecasting, outperforming traditional statistical and deep learning methods in over 86% of test cases. Its success highlights the power of large-scale pre-training for time series tasks, especially in providing equitable and accurate predictions for minority populations. Future research should explore multivariate models that incorporate external covariates such as economic indicators and policy changes to further enhance predictive capabilities.

Summary updated

BibTeX
@misc{akella2025comparative,
  author = {Aditya Akella and Jonathan Farah},
  title = {{Comparative Analysis of Time Series Foundation Models for Demographic Forecasting: Enhancing Predictive Accuracy in US Population Dynamics}},
  year = {2025},
  eprint = {2508.11680},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.11680},
}
Figure