ets4: October 21, 2025
Directly Relevant Papers
A curated list of recent papers on novel methods for forecasting economic time series.
In this Issue
- From News to Returns: A Granger-Causal Hypergraph Transformer on the Sphere
- Equity Market Price Changes Are Predictable: A Natural Science Approach
- Stochastic Volatility-in-mean VARs with Time-Varying Skewness
- Parsing the pulse: decomposing macroeconomic sentiment with LLMs
- Signature-Informed Transformer for Asset Allocation
- Adaptive Online Learning with LSTM Networks for Energy Price Prediction
- A three-step machine learning approach to predict market bubbles with financial news
- DP20727 Macroeconomic Forecasting and Machine Learning
Papers of Interest
Methodologically novel papers from other fields that could be adapted for economic forecasting.
In this Issue
- Multi-Agent Analysis of Off-Exchange Public Information for Cryptocurrency Market Trend Prediction
- An Adaptive Multi Agent Bitcoin Trading System
- Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing
- Aurora: Towards Universal Generative Multimodal Time Series Forecasting
From News to Returns: A Granger-Causal Hypergraph Transformer on the Sphere
- Authors: Anoushka Harit, Zhongtian Sun, Jongmin Yu
- Source: NEP-FOR (2025-10-21)
- Summary: The paper introduces the Causal Sphere Hypergraph Transformer (CSHT), a novel architecture designed for financial time-series forecasting that integrates Granger-causal hypergraph structures with Riemannian geometry and causally masked Transformer attention. This approach models the directional influence of financial news and sentiment on asset returns, providing robust generalization across market regimes and transparent attribution pathways from macroeconomic events to stock-level responses. The CSHT demonstrates superior performance in return prediction, regime classification, and asset ranking tasks, indicating its potential as a practical solution for forecasting under uncertainty.
Equity Market Price Changes Are Predictable: A Natural Science Approach
- Authors: Qingyuan Han
- Source: NEP-FOR (2025-10-21)
- Summary: The paper introduces the Extended Samuelson Model (ESM), which challenges the notion of equity market unpredictability by providing a framework that captures dynamic, causal processes in market behavior. ESM identifies market states and directional signals that can predict short-term reversals and longer-term trends, offering actionable insights for traders. This model’s ability to anticipate price changes even during breaking news events demonstrates its practical forecasting utility in financial markets.
Stochastic Volatility-in-mean VARs with Time-Varying Skewness
- Authors: Leonardo N. Ferreira, Haroon Mumtaz, Ana Skoblar
- Source: NEP-FOR (2025-10-21)
- Summary: This paper presents a Bayesian vector autoregression (BVAR) model that incorporates stochastic volatility-in-mean and time-varying skewness, allowing these factors to directly influence macroeconomic variables. The authors demonstrate that their model outperforms existing forecasting alternatives in a pseudo-real-time forecasting exercise, particularly in its ability to capture the effects of skewness shocks on output, inflation, and spreads. The findings emphasize the importance of including time-varying skewness for enhancing forecast accuracy and understanding macro-financial risks.
Parsing the pulse: decomposing macroeconomic sentiment with LLMs
- Authors: Byeungchun Kwon, Taejin Park, Phurichai Rungcharoenkitkul, Frank Smets
- Source: NEP-FOR (2025-10-21)
- Summary: This paper introduces a novel approach to forecasting macroeconomic indicators by utilizing Large Language Models (LLMs) to parse press narratives and extract sentiment indices related to growth and inflation. The authors demonstrate that these sentiment indices closely track traditional hard data and improve the forecasting performance of existing statistical models by incorporating information that is not captured by conventional datasets. The ability to decompose sentiment into its drivers—demand, supply, and structural forces—provides a deeper understanding of macroeconomic dynamics, enhancing predictive accuracy in real-time economic assessments.
Signature-Informed Transformer for Asset Allocation
- Authors: Yoontae Hwang, Stefan Zohren
- Source: NEP-FOR (2025-10-21)
- Summary: The paper introduces the Signature-Informed Transformer (SIT), a novel framework for asset allocation that directly optimizes a risk-aware financial objective, addressing the shortcomings of traditional deep-learning forecasters. Its innovative use of path signatures provides a rich geometric representation of asset dynamics, while the signature-augmented attention mechanism incorporates financial inductive biases, such as lead-lag effects. Evaluated on S&P 100 equity data, SIT significantly outperforms existing models, indicating its potential for enhancing predictive accuracy in financial decision-making.
Adaptive Online Learning with LSTM Networks for Energy Price Prediction
- Authors: Salih Salihoglu, Ibrahim Ahmed, Afshin Asadi
- Source: arXiv-csLG (2025-10-21)
- Summary: This paper presents a predictive model using Long Short-Term Memory (LSTM) networks to forecast day-ahead electricity prices in the California energy market. It introduces a novel custom loss function that combines Mean Absolute Error, Jensen-Shannon Divergence, and a smoothness penalty to enhance prediction accuracy. The model’s online learning approach allows it to adapt to new data incrementally, improving its performance and reducing prediction error, particularly during peak demand periods. The integration of various features, including historical prices and weather conditions, further strengthens its forecasting capabilities.
A three-step machine learning approach to predict market bubbles with financial news
- Authors: Abraham Atsiwo
- Source: arXiv-csLG (2025-10-21)
- Summary: This paper introduces a three-step machine learning framework for predicting market bubbles in the S&P 500 by integrating financial news sentiment with macroeconomic indicators. The novel aspect lies in its combination of textual data analysis through natural language processing and traditional econometric methods to enhance predictive accuracy. The empirical results demonstrate that the proposed ensemble learning approach significantly improves the prediction of bubble occurrences, offering valuable insights for stakeholders in managing financial risks.
DP20727 Macroeconomic Forecasting and Machine Learning
- Authors: Ta-Chung Chi, Ting-Han Fan, Raffaele Ghigliazza, Domenico Giannone, Zixuan (Kevin) Wang
- Source: CEPR_Papers (2025-10-11)
- Summary: This paper presents a novel approach to macroeconomic forecasting by integrating high-dimensional data with regularization techniques, rigorous out-of-sample validation, and the incorporation of nonlinearities. The authors demonstrate that regularization is crucial for managing model complexity, while the addition of nonlinearities offers limited gains in predictive accuracy. Their methodology aims to accurately predict the full conditional distribution of macroeconomic outcomes, providing a comprehensive view of macroeconomic risk in real time.
Multi-Agent Analysis of Off-Exchange Public Information for Cryptocurrency Market Trend Prediction
- Authors: Kairan Hong, Jinling Gan, Qiushi Tian, Yanglinxuan Guo, Rui Guo, Runnan Li
- Source: NEP-FOR (2025-10-21)
- Summary: This paper presents a multi-agent framework for predicting cryptocurrency market trends, addressing challenges such as volatility and news sensitivity. It introduces three key innovations: a news analysis system that quantifies market impact using large language models, a volatility-conditional fusion mechanism for combining news sentiment and technical indicators, and a distributed coordination architecture for real-time data processing. The comprehensive evaluation demonstrates significant improvements in prediction accuracy over existing methods, indicating a potential paradigm shift in financial machine learning applicable to quantitative trading and risk management.
- Reason for Interest: The innovative multi-agent framework and data processing techniques could be adapted to enhance forecasting models in economic time series, particularly in volatile markets like commodities or foreign exchange.
An Adaptive Multi Agent Bitcoin Trading System
- Authors: Aadi Singhi
- Source: NEP-CMP (2025-10-21)
- Summary: This paper introduces an Adaptive Multi Agent Bitcoin Trading System that leverages Large Language Models (LLMs) for predictive modeling in the highly volatile cryptocurrency market. The framework employs specialized agents for technical analysis and sentiment evaluation, and it incorporates a novel verbal feedback mechanism that allows the system to adapt its trading strategies over time without the need for parameter updates. Back-testing results indicate significant outperformance compared to traditional buy-and-hold strategies, showcasing the potential for improved forecasting in financial trading through innovative use of LLMs.
- Reason for Interest: The adaptive feedback mechanism and multi-agent approach could be applied to economic forecasting by enhancing predictive models for volatile economic indicators, allowing for real-time adjustments based on market sentiment and other dynamic factors.
Explainable Heterogeneous Anomaly Detection in Financial Networks via Adaptive Expert Routing
- Authors: Zan Li, Rui Fan
- Source: arXiv-csLG (2025-10-21)
- Summary: This paper presents a novel approach to anomaly detection in financial networks that incorporates adaptive graph learning and specialized expert networks, allowing for the identification of specific types of financial anomalies with a lead time of 3.8 days. The framework utilizes BiLSTM with self-attention to capture multi-scale temporal dependencies and integrates temporal and spatial information through cross-modal attention. The method’s interpretability is architecturally embedded, enabling automatic identification of anomaly mechanisms without labeled supervision, which could significantly enhance predictive capabilities in economic contexts.
- Reason for Interest: The innovative anomaly detection methods and adaptive graph learning could be applied to economic forecasting, particularly in identifying and predicting market shocks or regime changes.
Aurora: Towards Universal Generative Multimodal Time Series Forecasting
- Authors: Xingjian Wu, Jianxin Jin, Wanghui Qiu, Peng Chen, Yang Shu, Bin Yang, Chenjuan Guo
- Source: arXiv-csLG (2025-10-21)
- Summary: The paper introduces Aurora, a Multimodal Time Series Foundation Model designed for effective time series forecasting across different domains. It innovatively incorporates multimodal inputs, such as text and images, to enhance the extraction of domain-specific knowledge, which is crucial for accurate predictions. The model supports zero-shot inference, allowing it to generalize across domains without needing retraining, and employs a novel Prototype-Guided Flow Matching technique for generative probabilistic forecasting. Comprehensive experiments demonstrate its state-of-the-art performance in both unimodal and multimodal scenarios, indicating significant advancements in forecasting methodologies.
- Reason for Interest: The innovative multimodal approach and zero-shot inference capabilities could be adapted to improve economic forecasting by integrating diverse data sources, such as news articles and market sentiment indicators.