Probabilistic Forecasting of Non-Stationary Epidemic Dynamics

24 concepts22 papers

Traditional epidemiological models often operate under an assumption of clean periodicity: if a respiratory virus peaks in January of this year, it ought to follow a similar trajectory next year. In reality, real-world disease tracking is plagued by extreme volatility. The timing, peak magnitude, and spatial spread of outbreaks vary drastically from one season to the next.

This high inter-annual variability creates a noisy, non-stationary environment where historical benchmarks quickly lose their utility. When a novel pandemic strain or variant emerges, the challenge scales exponentially due to a lack of real-time data on transmission dynamics and severity parameters. To overcome these limitations, modern epidemiological forecasting has shifted away from rigid, deterministic models toward flexible, data-driven probabilistic frameworks and collaborative ensemble methods.

The Challenge of Non-Stationarity and Volatility

The core obstacle in epidemiological time-series forecasting is non-stationarity—a property where the statistical rules governing disease transmission change over time. As highlighted by McGowan et al. and Farrow et al., the inherent volatility of viral circulation makes it exceptionally difficult to predict the exact onset or peak of an outbreak.

This non-stationarity manifests in several ways:

  • Inter-annual shifts: The dominant viral strains and environmental triggers evolve, causing the strength and timing of seasonal patterns to shift across years (Lau et al.; Albrecht et al.).
  • Abrupt phase transitions: During an epidemic, a population transitions rapidly through phases of increasing, peaking, and decreasing case counts. Models that perform well during quiet periods frequently fail to maintain accuracy during these sudden, dynamic shifts (Lopez et al.).
  • Novel pathogen uncertainty: When a entirely new strain emerges, early data is obscured by reporting delays and case-detection biases, leaving researchers with an incomplete picture of the pathogen's natural history (Nsoesie et al.; Lv et al.).

Whether tracking biological viruses or modeling the propagation speed of emerging computer malware strains (Yom-Tov et al.), standard periodic models struggle to adapt to these jagged fluctuations.

Flexible Data-Driven Learning

To capture non-stationary behavior without forcing the data into a restrictive mathematical box, researchers utilize non-parametric statistical frameworks and deep learning.

Gaussian Processes for Adaptive Uncertainty

Gaussian Processes (GPs) offer a powerful alternative to traditional mechanistic modeling by treating the forecasting problem probabilistically. Instead of assuming a fixed functional form for how a disease spreads, a GP models a distribution over all possible transmission curves.

Senanayake et al. demonstrated this by using Variational Gaussian Process Regression paired with a custom "averaging kernel" to capture complex geographical dependencies and seasonal non-stationarity across massive spatio-temporal datasets. Because GPs inherently quantify uncertainty, they adapt gracefully to unusual seasons. As Zimmer et al. showed, training specialized GP models for specific temporal targets yields superior probabilistic accuracy (log-scores) because the model can widen its uncertainty bounds precisely when encountering volatile population behaviors.

Deep Time-Series Decomposition

Deep learning architectures approach non-stationarity by breaking complex epidemic curves into simpler, isolated components. For instance, Zheng designed a hybrid framework that decomposes volatile pandemic time-series data before passing the sub-signals into optimized recurrent networks (LSTM and Elman). Similarly, Yang leveraged a bidirectional LSTM optimized via a whale optimization algorithm to maintain long-term accuracy amidst rapidly changing epidemic dynamics. By separating the underlying trend from sudden, short-term noise, these models prevent localized spikes from distorting long-term projections.

Why Multi-Model Ensembles Work

Even the most advanced individual model can fail when a pandemic undergoes a structural shift—such as a resurgence driven by a new variant or a sudden change in public policy (Truelove et al.; the US COVID-19 Scenario Modeling Hub evaluation). Individual models are prone to overconfidence because they are bound to their specific mathematical assumptions.

The solution is collaborative, multi-model ensembling. By aggregating predictions from diverse teams, ensembles hedge against the structural blind spots of any single approach.

Individual Model Forecasts                 Aggregated Ensemble
┌─────────────────────────┐               ┌─────────────────────────┐
│  Model A (Mechanistic)  ├─┐             │                         │
└─────────────────────────┘ │             │                         │
┌─────────────────────────┐ ├─► Mean/Median│   Robust, Smoothed     │
│  Model B (Deep Learning)├─┼──► Ensemble ├──► Probabilistic Forecast│
└─────────────────────────┘ │             │  (Reduces Overfusion)   │
┌─────────────────────────┐ │             │                         │
│  Model C (Statistical)  ├─┘             │                         │
└─────────────────────────┘               └─────────────────────────┘

The statistical intuition behind this performance is robust:

  1. Error Cancellation: If Model A overestimates transmission due to rigid herd-immunity assumptions, and Model B underestimates it by over-indexing on recent data, an unweighted mean ensemble averages out these uncorrelated errors (McGowan et al.).
  2. Density Combination: Methods like the "delta density" approach developed by Farrow et al. combine full week-ahead probability distributions rather than single point estimates, preserving vital uncertainty metrics.
  3. Outlier Insensitivity: In highly volatile environments, specific models can experience catastrophic failures. Sherratt et al. demonstrated that median-based ensembles consistently outperform both individual models and simple averages across diverse nations because the median operation is naturally resilient to extreme, erroneous forecast outliers.

Overcoming Data Sparsity via Information Borrowing

When a novel pathogen or variant first emerges, local data is often too sparse or noisy to train reliable predictive models (Lee et al.). In these early stages, external shocks—such as a sudden lifestyle shift or a new intervention—can completely alter expected disease trends (Li et al.).

To generate reliable real-time predictions before local baseline data stabilizes, researchers use Bayesian hierarchical modeling to borrow information across different regions.

Lee et al. demonstrated that by treating individual nations as distinct but mathematically related realizations of a shared global epidemic process, the growth curves observed in countries further along the outbreak timeline can act as a structured prior for countries just beginning to see cases. This shared statistical strength constrains the model's parameters, preventing it from over-fitting to early local noise while still allowing for regional customization as more data arrives.

Key Foundational Work

A systematic review of studies on forecasting the dynamics of influenza outbreaks (2013) — contributed 3 concept elements to this direction.
Flexible Modeling of Epidemics with an Empirical Bayes Framework (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_predicting_novel_pandemic_strains_8d1688cf,
  title = {predicting novel pandemic strains},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predicting-novel-pandemic-strains/rdr_8d1688cf54f240549d5b43b3605011b6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_predicting_novel_pandemic_strains_8d1688cf_v1,
  title = {predicting novel pandemic strains},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predicting-novel-pandemic-strains/rdr_8d1688cf54f240549d5b43b3605011b6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}

See something wrong? Report a mistake