Forecasting Geopolitical Events with a Sparse Temporal Fusion Transformer and Gaussian Process Hybrid: A Case Study in Middle Eastern and U.S. Conflict Dynamics

Hsin-Hsiung Huang, Hayden Hampton · 2025

Read Forecasting Geopolitical Events with a Sparse Temporal Fusion Transformer and Gaussian Process Hybrid: A Case Study in Middle Eastern and U.S. Conflict Dynamics by Hsin-Hsiung Huang, Hayden Hampton (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-20935.txt.

Introduction

Hsin-Hsiung Huang and Hayden Hampton forecast geopolitical conflict with STFT-VNNGP, a transformer-Gaussian process hybrid built for spiky event data. Their paper from twenty twenty-five uses GDELT records, whose sparsity and burstiness wreck standard deep forecasters at long horizons. Stage one runs a sparse temporal fusion transformer to produce multi-quantile forecasts. Stage two feeds those quantiles to a variational nearest neighbor Gaussian process.

The process layer smooths spatiotemporally and quantifies uncertainty on principle. The architecture won a twenty twenty-three threat detection competition. In Middle East and United States case studies, it beats a lone transformer, especially on timing and size of bursty episodes far ahead. Code and workflows are public for reproduction.

Catalog record

CREATORHsin-Hsiung Huang, Hayden Hampton
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-20935

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Forecasting Geopolitical Events with a Sparse Temporal Fusion Transformer and Gaussian Process Hybrid: A Case Study in Middle Eastern and U.S. Conflict Dynamics Hsin-Hsiung Huang, Hayden Hampton Abstract Forecasting geopolitical conflict from data sources like the Global Database of Events, Language, and Tone (GDELT) is a critical challenge for national security. The inherent sparsity, burstiness, and overdispersion of such data cause standard deep learning models, including the Temporal Fusion Transformer (TFT), to produce unreliable long-horizon predictions. We introduce STFT-VNNGP, a hybrid architecture that won the 2023 Algorithms for Threat Detection (ATD) competition by overcoming these limitations. Designed to bridge this gap, our model employs a two-stage process: first, a TFT captures complex temporal dynamics to generate multi-quantile forecasts. These quantiles then serve as informed inputs for a Variational Nearest Neighbor Gaussian Process (VNNGP), which performs principled spatiotemporal smoothing and uncertainty quantification. In a case study forecasting conflict dynamics in the Middle East and the U.S., STFT-VNNGP consistently outperforms a standalone TFT, showing a superior ability to predict the timing and magnitude of bursty event periods, particularly at long-range horizons. This work offers a robust framework for generating more reliable and actionable intelligence from challenging event data, with all code and workflows made publicly available to ensure reproducibility. Keywords: Conflict Prediction; Deep Learning; Overdispersion; Spatiotemporal Smoothing; Uncertainty Quantification. 1 Introduction The accurate and timely forecasting of geopolitical unrest is critical for diplomatic decision-making, humanitarian planning, and national security risk assessment ( Ward et al. 2010 ) . This complex forecasting problem was the focus of the 2023 Algorithms for Threat Detection (ATD) program, a challenge jointly organized by the U.S. National Science Foundation and the National Geospatial-Intelligence Agency. The ATD program is known for posing difficult, large-scale forecasting challenges on real-world data, from reconstructing vessel trajectories using sparse maritime sensors ( Chen and Huang 2023 ) to detecting anomalies in traffic flows with extensive missing data ( He et al. 2023 ; Yu et al. 2024 ) . The 2023 challenge, which this paper addresses, centered on producing multi-horizon forecasts of weekly event counts for 20 CAMEO event classes across thousands of spatio-temporal series extracted from the Global Database of Events, Language, and Tone (GDELT) ( Leetaru and Schrodt 2013 ) . Forecasting on this data presents a formidable set of statistical challenges that render standard time series models inadequate. First, the task requires multi-horizon forecasting —predicting event counts not just one week ahead, but for several weeks into the future. This is notoriously difficult, as predictive uncertainty grows rapidly with the forecast horizon. Iterative methods suffer from error accumulation, while direct methods require training separate models for each horizon, increasing complexity and potentially ignoring temporal dependencies between horizons ( Lim et al. 2021 ) . Second, the event data are extremely sparse and zero-inflated , with over 90% of weekly observations being exact zeros. This high degree of sparsity violates the assumptions of classical models like ARIMA and poses significant optimization difficulties for complex, gradient-based deep learning architectures, which can struggle to learn meaningful patterns from infrequent positive signals ( Brandt and Williams 2001 ) . Third, the non-zero counts are bursty : long periods of quiescence are punctuated by abrupt, large-magnitude spikes in activity. This burstiness is a well-documented feature of conflict data and social unrest, often following heavy-tailed distributions ( Bohorquez et al. 2009 ) . Global models, including modern deep learning architectures like the Temporal Fusion Transformer (TFT) ( Lim et al. 2021 ) , are trained to minimize average error across many series. Consequently, they tend to smooth over these localized, high-impact bursts, leading to systematic under-prediction during the very critical events that early-warning systems are designed to detect. This paper introduces a hierarchical hybrid model designed to address this trifecta of challenges. Our approach, the Spatio-Temporal Fusion Transformer with a Variational Nearest-Neighbor Gaussian Process (STFT–VNNGP), decomposes the event rate into a global, non-linear trend and a local, spatio-temporal residual process. The global component is captured by a TFT, which leverages rich feature representations and attention mechanisms to model complex temporal dynamics shared across all series. The local component, which models site-specific deviations, is captured by a Variational Nearest-Neighbor Gaussian Process (VNNGP) ( Wu et al. 2022 ) . This GP layer imposes spatial coherence and provides the flexibility to model the abrupt, bursty dynamics that the global model fails to capture. By combining the global predictive power of the TFT with the local adaptability and principled uncertainty quantification of the GP, our model provides a more robust and accurate tool for operational early-warning systems. We demonstrate through empirical tests on the ATD benchmark that our hybrid approach significantly reduces forecast error compared to a standalone TFT. 2 GDELT Event Data 2.1 Primary Evaluation Dataset: Middle East (Weekly) Our primary dataset consists of weekly event aggregates from the GDELT 2.0 Event Database, which monitors global news media to codify “who did what to whom” ( Leetaru and Schrodt 2013 ) . The data cover the period from January 2014 to December 2023. We focus on five states within the United States Central Command’s area of responsibility—Iraq, Syria, Lebanon, Jordan, and Israel—which exhibit complex internal dynamics and significant cross-border spillover effects. Event records are geolocated and aggregated to administrative level-1 units (e.g., provinces or governorates), yielding 97 distinct spatial locations. Each location’s time series is further partitioned into 20 mutually exclusive event types based on the Conflict and Mediation Event Observations ( CAMEO ) taxonomy ( Schrodt et al. 2012 ) . CAMEO provides a detailed schema for classifying interactions between political actors, ranging from peaceful diplomatic cooperation (e.g., code 04: Consult ) to conventional military action (e.g., code 19: Fight ). The resulting data tensor has dimensions 97\text{ (provinces)}\times 20\text{ (event types)}\times 522\text{ (weeks)} , forming 1,940 unique series. The statistical properties of this dataset underscore the modeling challenges. The overall sparsity ratio, defined as the proportion of zero-valued entries, is 0.927. Conditional on an event count being non-zero, the data exhibit extreme positive skew: the median count is 3, while the 99th percentile exceeds 120. This confirms the pronounced burstiness that motivates our hybrid modeling approach, a characteristic that makes accurate point forecasting exceptionally difficult. 2.2 Secondary Validation Dataset: Global (Daily) To assess the robustness and external validity of our model, we use a secondary dataset derived from GDELT 1.0 ( Leetaru and Schrodt 2013 ) . This panel records daily events from January 2007 to December 2009 in which the United States is coded as either Actor 1 or Actor 2. While this dataset is global rather than region-specific and has a daily rather than weekly resolution, it shares the same defining statistical properties: extreme sparsity (94% zeros) and heavy-tailed, bursty counts. The consistent performance of our model across both datasets strengthens the generalizability of our conclusions. (a) Diplomatic Cooperation (b) Military Posture (c) Protest Events (d) Reduce Relations Figure 1 : Weekly event counts for four CAMEO event types across Israel’s eight ADMIN1 regions, illustrating data sparsity and burstiness. 3 Methodology We propose the Sparse Temporal Fusion Transformer–Variational Nearest Neighbor Gaussian Process (STFT–VNNGP), a multi-stage hybrid framework designed for forecasting challenging spatio-temporal event counts. Geopolitical data, such as GDELT, exhibit distinct statistical behaviors

Read the complete text