Accurate financial time series forecasting hinges on effectively integrating both numerical price data and external textual signals, such as news events. While Time-LLM provides a cross-modal time-series prediction mechanism for price sequences, our main contribution focuses on a novel method to fuse these time-series embeddings with learned representations of news text within a large language model (LLM) framework. In contrast to traditional approaches, often relying on text-only training or simple numeric similarity for retrieval, our method leverages the LLM’s capacity to capture intricate correlations between language-based context and price fluctuations. To facilitate this, we introduce aligned datasets that pair daily (or periodic) financial news with the corresponding market data, ensuring temporal consistency. Our model performs token-level concatenation of textual embeddings extracted from a domain-specific financial language model and patched numerical embeddings, followed by frozen LLaMA self-attention in a shared embedding space. By jointly encoding textual and numerical features, the model captures interactions between historical market patterns and temporally aligned textual context. We evaluate the approach on three datasets spanning gold prices, foreign exchange rates, and Bitcoin, which include structured macroeconomic-calendar descriptions and news-derived textual summaries. Experimental results show that the incorporation of aligned textual information improves the accuracy of forecasting, underscoring the importance of robust integration of news data in financial prediction tasks.
Since Indonesia’s first COVID-19 outbreak, central and regional governments have used individual-level health data to inform policy, creating privacy concerns for data sharing. This study evaluates Bayesian network (BN) methods for generating synthetic COVID-19 surveillance data from Jakarta and compares four BN algorithms with parametric and non-parametric Synthpop and two sparse-autoencoder (SAE) configurations. Utility was assessed through distributional similarity, logistic-regression inference, and machine-learning prediction; privacy was assessed using sdcMicro record distinctiveness and TCAP attribute-disclosure risk. Across ten synthetic replicates, BN methods preserved the observed distributions more closely than SAE and showed TCAP values close to parametric Synthpop, although SAE yielded lower sdcMicro risk. BN and Synthpop reproduced several logistic-regression coefficient patterns, but some variables changed in magnitude or relation classification. In predictive analyses, BN and non-parametric Synthpop generally produced performance closer to the real-data baseline than parametric Synthpop or SAE. These findings indicate a method- and metric-specific privacy–utility trade-off and support using synthetic data only after analysis-specific validation.
Computational propaganda campaigns came into the limelight after the 2016 US elections and since then the research has predominately focused on foreign influence operations across the US, UK, and EU. Computational propaganda campaigns in the Global South remain understudied. Therefore, we shift our focus towards Pakistan and examine computational propaganda campaigns by domestic political actors. Utilizing Facebook digital trace data, we track 400 hashtag campaigns by domestic political actors to analyze: (1) the dissemination patterns of political information, (2) the role bot and spam accounts play in the dissemination of political information, (3) the nature of hashtag campaigns, and (4) the key initiators and amplifiers during the hashtag campaigns. Our research contributions include: (1) we validate the findings in the previous literature regarding presence of bots and increased coordinated activity during the emergence of political events and show that the most central bridging positions within these campaigns are held by organized spam accounts. (2) Our analysis of multi-lingual hashtag campaigns uncovers that language and political context play a pivot role in dissemination of political information. (3) Tracing campaigns across several years, we find that automated and inauthentic activity intensifies around electoral contests, rising to roughly two-fifths of monthly posts by the 2024 general election, when spam accounts came to outnumber verified institutional actors. Together, these findings characterize computational propaganda in Pakistan as a domestically driven, linguistically stratified, and as an issue/event-driven phenomenon.
Social media has become a relevant platform to communicate with a vast audience, enabling users to spread information or influence discussions. As such, it has garnered the attention of governments and other organizations, enabling them to engage with their audiences directly and bypass traditional channels that might disrupt their message. Given their influence on shaping public perception, understanding and analyzing how they communicate provides insights into their political strategy and their outward perception. We have conducted a study based on a collection of tweets sent between January 2020 and March 2021 from diplomats in four different regions: Russia, China, the European Union, and the United States. We have analyzed the content produced and the properties of the following network connecting these users. Our findings show a clear structural separation in the following network by region, which may reflect differences in their diplomatic strategies. The content analysis revealed the use of propagandistic content during key events, suggesting that observed engagement between diplomats from distinct regions typically follows existing structural paths. With this work, we aim to fill a gap in digital diplomacy by focusing on the behavior of diplomats of global powers across three continents.
This paper develops a deep machine learning framework for analyzing the role of topological heterogeneities in multivariate time-series data. Topological heterogeneities—global invariants of the shape of the data at each point in time—are extracted and quantified using topological data analysis (TDA) in the form of persistence diagrams, which record the occurrence and stability of these invariants. Rather than reducing persistence diagrams to functional summaries before learning, as is standard practice, the framework ingests each diagram as a structured, homology-aware, time-aligned signal that conditions a supervised deep classifier of covariate sequences. Applied to international-conflict data on militarized interstate disputes across ten model architectures under 5-fold cross-validation, incorporating topological heterogeneity improves classification accuracy and F1 over matched baselines under pooled significance tests, while ranking and calibration metrics are largely unaffected and confidence on incorrect predictions rises slightly once probabilities are temperature-scaled. A feature-importance analysis shows that this topological conditioning does not reorder the leading covariates compared to non-topologically conditioned base model. It does however make their importance more consistent across architectures, which is the signature of a conditioning rather than a determining signal. The results indicate that global topological configuration carries an irreducibly systemic component of the data that unit-level, covariate-based modeling cannot recover, and that it is a largely untapped source of information for time-series analysis in the social sciences.
