|
on Big Data |
| By: | Alice Treesa M; Dr. Arpita Choudhary (Assistant Professor, Madras School of Economics, Chennai, India.) |
| Abstract: | Reliable energy consumption forecasting in the short term is essential for improving building operations efficiency and creating sustainable energy consumption plans. The authors of this study evaluate the forecasting performance of machine learning and deep learning methods which use climate data and time data to predict energy usage at hourly intervals. The study used Linear Regression, Decision Trees, Random Forest, XGBoost and Long Short-Term Memory as comparison methods to assess performance in the same context. The study demonstrated that energy consumption forecasting accuracy depends more on selected features than on the model's complexity. The study found that LSTM model learning capacity remained stable while Random Forest model performance showed superior results in dealing with non-linear features that had temporal attributes. |
| Keywords: | Energy Consumption Prediction, Machine Learning, Ensemble Models, LSTM Model, Feature Engineering, Sustainable Energy Managementsemantics, Neural architectures |
| JEL: | Q47 C53 C45 C38 L94 Q41 |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:mad:wpaper:2026-301 |
| By: | Abdukakhkhor Abdurakhmonov (Central Bank of Uzbekistan) |
| Abstract: | This paper provides the first systematic assessment of machine learning methods for macroeconomic forecasting in Uzbekistan. Using a comprehensive dataset of more than 170 indicators, we forecast CPI inflation and GDP growth with nine machine learning models and compare them against three traditional benchmarks (ARIMA, VAR, and BVAR). For both targets, the relative performance of machine learning improves as the forecast horizon increases. For inflation, machine learning provides clear and growing gains as the horizon increases, and a simple equal-weighted ensemble of the machine learning models is the most accurate approach overall, achieving the lowest forecast error at nearly every horizon. For GDP growth, by contrast, the traditional benchmarks (ARIMA in particular) remain the most accurate across most horizons, although regularized linear and dimension-reduction machine learning methods are competitive at short horizons. Tree-based models struggle to forecast GDP when growth exceeds the range observed during training because they cannot extrapolate beyond the training data. This limitation is particularly relevant in Uzbekistan's rapidly changing economy, where rapid economic growth in 2024-2025 pushed the level of GDP beyond the range observed in the training sample. We show that forecasting stationary transformations of the target largely removes this weakness. Overall, the findings suggest that machine learning is best used to complement rather than replace the existing forecasting toolkit. It improves the accuracy of medium-term inflation forecasts, whereas traditional models remain more accurate for forecasting GDP. |
| Keywords: | Machine Learning; Macroeconomic Forecasting; Ination; GDP Growth; Model Evaluation and Selection; Uzbekistan |
| JEL: | C22 C45 C53 E31 E37 E52 |
| Date: | 2026–08–03 |
| URL: | https://d.repec.org/n?u=RePEc:gii:giihei:heidwp19-2026 |
| By: | Lönn, Gabriel Elias (University of Oslo); Schutte, Sebastian |
| Abstract: | Text-as-data methods aim to extract quantitative information from natural language. Traditionally, this was accomplished by using separate steps in each analysis, such as part-of-speech tagging, sentiment analysis, and named entity recognition. Modern Large Language Models (LLMs) hold the promise of drastically simplifying this process, as they can be flexibly instructed to extract specific information. However, transferring data to LLM providers at scale and receiving directly usable data back has previously required application-specific programming. In contrast, the rapidcodeR package offers a highly flexible approach to coding quantitative data from text at maximum speed and minimal cost. Here, we present an example of using the package to visualize international relations, directly coded from Russian UN speeches between 1946 and 2024. The results correspond well to formal alliance structures. Similar use cases can involve coding of archival information, researching social media discourse, and extracting event data from news sources. |
| Date: | 2026–07–22 |
| URL: | https://d.repec.org/n?u=RePEc:osf:socarx:q4zd8_v1 |
| By: | Rauh, C. |
| Abstract: | This paper proposes a methodology to extract information about regional economic conditions from newspaper text in real time. The approach relies on large-scale collections of news articles that are summarized using unsupervised machine learning to generate topics capturing recurring themes in economic reporting. Because the method uses the full corpus of regional news and avoids restrictive keyword selection, it minimizes human judgment and allows the data to reveal economically relevant patterns in news coverage. I apply the methodology to Canada, a large and economically diverse country, and show that the resulting topics contain information about fluctuations in economic indicators such as manufacturing activity and unemployment at both the national and provincial levels. The results are robust to alternative choices of the number of topics. A composite index constructed from the topic measures provides a summary indicator of economic information contained in the news. While some topics display similar associations with economic outcomes across provinces, others capture region-specific developments, highlighting the ability of the approach to uncover geographically heterogeneous economic signals in news data. |
| Keywords: | Machine Learning, Latent Dirichlet Allocation, Newspaper Text, Economic Uncertainty, Topic Model, Canada |
| JEL: | D80 E66 C55 |
| Date: | 2026–06–26 |
| URL: | https://d.repec.org/n?u=RePEc:cam:camdae:2657 |
| By: | Araujo, Douglas; Bokan, Nikola; Comazzi, Fabio; Lenza, Michele |
| Abstract: | Word embeddings are vectors of real numbers associated with words, designed to capture semantic and syntactic similarity between the words in a corpus of text. We estimate the word embeddings of the European Central Bank's introductory statements at monetary policy press conferences by using a simple natural language processing model (Word2Vec), only based on the information and model parameters available as of each press conference. We show that a measure based on such embeddings contributes to improve core inflation forecasts multiple quarters ahead. Other common textual analysis techniques, such as dictionary-based metrics or sentiment metrics do not obtain the same results. The information contained in the embeddings remains valuable for out-of-sample forecasting even after controlling for the central bank inflation forecasts, which are an important input for the introductory statements. |
| Keywords: | Inflation |
| JEL: | E31 E37 E58 |
| Date: | 2024–12 |
| URL: | https://d.repec.org/n?u=RePEc:cpr:ceprdp:19784 |
| By: | Fernández-Villaverde, Jesús; Nuño, Galo; Perla, Jesse |
| Abstract: | We argue that deep learning provides a promising avenue for taming the curse of dimensionality in quantitative economics. We begin by exploring the unique challenges posed by solving dynamic equilibrium models, especially the feedback loop between individual agents' decisions and the aggregate consistency conditions required by equilibrium. Following this, we introduce deep neural networks and demonstrate their application by solving the stochastic neoclassical growth model. Next, we compare deep neural networks with traditional solution methods in quantitative economics. We conclude with a survey of neural network applications in quantitative economics and offer reasons for cautious optimism. |
| Keywords: | Deep learning |
| JEL: | C61 C63 E27 |
| Date: | 2024–11 |
| URL: | https://d.repec.org/n?u=RePEc:cpr:ceprdp:19636 |
| By: | Junyu Chen; Tom Boot; Lingwei Kong; Weining Wang |
| Abstract: | Conditional Value-at-Risk (CoVaR) quantifies systemic financial risk by measuring the loss quantile of one asset, conditional on another asset experiencing distress. We develop a Transformer-based methodology that integrates financial news articles directly with market data to improve CoVaR estimates. Unlike approaches that use predefined sentiment scores, our method incorporates raw text embeddings generated by a large language model (LLM). We prove explicit error bounds for our Transformer CoVaR estimator, showing that accurate CoVaR learning is possible even with small datasets. Using U.S. market returns and Reuters news items from 2006–2013, our out-of-sample results show that textual information impacts the CoVaR forecasts. With better predictive performance, we identify a pronounced negative dip during market stress periods across several equity assets when comparing the Transformer-based CoVaR to both the CoVaR without text and the CoVaR using traditional sentiment measures. Our results show that textual data can be used to effectively model systemic risk without requiring prohibitively large data sets. |
| Date: | 2026–01–30 |
| URL: | https://d.repec.org/n?u=RePEc:bri:uobdis:26/840 |
| By: | Kellner, Domenic; Lang, Jan Hannes; Rusnák, Marek; Nagy, Lukas Joseph |
| Abstract: | Financial stability risks consist of two distinct components: vulnerabilities and possible trigger events. While there has been considerable progress regarding the measurement of vulnerabilities, the assessment of possible trigger events remains largely qualitative. To fill this gap, we employ Large Language Models to extract information about the Severity and Probability Of potential Trigger events (SPOT) from a large dataset of financial news articles over the period2005 – 2026. The SPOT indicator increases ahead of major historical trigger events, correctly identifies trigger sources, and helps to improve forward looking model estimates of downside risks to the economy. The results indicate that the use of AI-based signal extraction from text can be a promising avenue to improve the monitoring of financial stability risks. JEL Classification: C55, C88, E32, E44, G01 |
| Keywords: | artificial intelligence, crisis indicators, financial stability, growth-at-risk |
| Date: | 2026–07 |
| URL: | https://d.repec.org/n?u=RePEc:ecb:ecbwps:20263262 |
| By: | Prashanth BS; Manoj Kumar; Ariful Hoque; Nasser Al Muraqab; Immanuel Azaad Moonesar; Udo Christian Braendle; Ananth Rao |
| Abstract: | The development of online banking has brought about an increase in fraudulent operations, which is a major problem for banks. This study delves into the urgent requirement for interpretable, scalable, and top-notch fraud detection systems by using TabNet, an adaptable deep learning framework, on a Kaggle dataset consisting of actual bank transactions in India. Maximizing operational risk management by improving the accuracy of transaction anomaly detection and ensuring regulatory compliance through transparent models is the goal. We utilize a supervised learning pipeline that incorporates the Synthetic Minority Oversampling Technique (SMOTE) to ensure that classes are balanced. Subsequently, we conduct thorough exploratory data analysis (EDA) to identify patterns of fraud, both during specific times and across behaviors. On this dataset, five different deep learning architectures are tested: DNN, GRU, LSTM, CNN1D, and TabNet. Assessment of predictive performance was carried out using a 3-fold cross-validation framework. With a ROC-AUC of 0.9739 and an accuracy of 97.39 %, TabNet considerably outperformed the competition. The method of sparse feature selection used improved interpretability, generalized better on tabular data, and produced fewer false positives and negatives. Critical insights for operational fraud detection systems and a contribution to the broader literature on explainable AI (XAI) in financial decision-making are offered by the findings. Goals 8 and 16 of the Sustainable Development Agenda are supported by this study, which promotes inclusive economic growth and institutional transparency. Supporting strong, policy-compliant, and interpretable decision-support systems, it also offers practical use for real-time implementation in banking infrastructure. |
| Date: | 2026–07 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2607.18616 |
| By: | Luc Hazenoot; Zhaochun Ren; Amirhossein Zohrehvand |
| Abstract: | Existing measures of how much a text is about a concept read the surface of the text: dictionary word shares, topic proportions, embedding similarities. They score the words a text uses, not the judgment a reader forms about it. Recent work has shown that a gap exists in what Large Language Models (LLMs) know internally versus what they express in their response. This paper asks whether that internal knowledge, read by monitoring the activations of frozen, out-of-the-box LLMs, can stand in for task-specific fine-tuning when measuring concept content, and which extraction method reads it best. We extract such measures via the Recursive Feature Machine (RFM) algorithm and via linear probing, and compare these against an embedding baseline, surface baselines, and the same model's own answer to the question. We demonstrate the approach on financial text, a domain studied extensively and served by established annotated resources, using a human-annotated Environmental, Social and Governance (ESG) dataset. The best linear probe comes within 0.6 percentage points of a fine-tuned domain classifier's accuracy without any task-specific fine-tuning, and outscores the same model's own answer to the question in eleven of twelve comparisons, so the activations carry concept content the response does not report. The simple probe consistently beats the RFM concept vectors, which in turn provide what classification alone does not: a continuous score intended to reflect how strongly a concept is present in a text, whose validation awaits graded labels. |
| Date: | 2026–08 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2608.07208 |
| By: | Galasso, Vincenzo; Nannicini, Tommaso; Nozza, Debora |
| Abstract: | Understanding individuals’ beliefs, preferences, and motivations is essential in social sciences. Recent technological advancements—notably, large language models (LLMs) for analyzing open-ended responses and the diffusion of voice messaging— have the potential to significantly enhance our ability to elicit these dimensions. This study investigates the differences between oral and written responses to open-ended survey questions. Through a series of randomized controlled trials across three surveys (focused on AI, public policy, and international relations), we assigned respondents to answer either by audio or text. Respondents who provided audio answers gave longer, though lexically simpler, responses compared to those who typed. By leveraging LLMs, we evaluated answer informativeness and found that oral responses differ in both quantity and quality, offering more information and containing more personal experiences than written responses. These findings suggest that oral responses to open-ended questions can capture richer, more personal insights, presenting a valuable method for understanding individual reasoning. |
| Keywords: | Survey design; Open-ended questions; Large Language Models; beliefs |
| JEL: | C83 D83 |
| Date: | 2024–12 |
| URL: | https://d.repec.org/n?u=RePEc:cpr:ceprdp:19749 |
| By: | Braghieri, Luca; Eichmeyer, Sarah; Levy, Ro'ee; Mobius, Markus; Steinhardt, Jacob; Zhong, Ruiqi |
| Abstract: | There is widespread concern that the social media ecosystem drives users to engage with like-minded news articles, thereby fostering polarization in news consumption. Methodological limitations in estimating slant at the article level have made evaluating these claims difficult. We use data on the near universe (ca. 1 million) of hard news articles published online by the top 100 U.S. news outlets in 2019, together with recent advances in natural language processing, to obtain a content-based measure of slant at the article level. Our main finding is that the degree of polarization in news consumption on social media is arguably high. Specifically, the mean slant difference between articles consumed by conservative and liberal users on Facebook is 1.5 times the ideological distance between the average New York Times and Foxnews.com article. We also show that: i) the majority (65%) of the variance in slant across articles arises within outlets, rather than across outlets, highlighting the importance of measuring slant at the article rather than the outlet level. ii) Most news produced is centrist, but the tails of the slant distribution are thick and there is substantial variation in slant across news type and topic. iii) Extreme content is much more likely to be shared widely on Facebook than moderate content. iv) There is substantial pro-attitudinal news consumption on Facebook even within the same outlet. v) Polarization in news exposure can account for the majority of polarization in news consumption on Facebook. |
| Keywords: | News |
| JEL: | D72 L82 L86 Z13 |
| Date: | 2024–12 |
| URL: | https://d.repec.org/n?u=RePEc:cpr:ceprdp:19807 |
| By: | De Fiore, Fiorella; Maurin, Alexis; Mijakovic, Andrej; Sandri, Damiano |
| Abstract: | We analyse the media’s role in channelling information about the Fed’s monetary policy stance to the public. Using LLMs, we find a tight correspondence between FOMC communication and media coverage, although with significant variation over time. The communication pass-through weakened during the ZLB period and improved with the introduction of press conferences, which now exert strong influence on the media. Media coverage effects households' inflation expectations, particularly when inflation is high and volatile, while we do not detect a direct impact of FOMC communication. This underscores the media’s crucial function in channelling central banks’ communication to the public. |
| Keywords: | Central bank communication |
| JEL: | E50 E52 E58 |
| Date: | 2024–12 |
| URL: | https://d.repec.org/n?u=RePEc:cpr:ceprdp:19748 |
| By: | Tamkin Nuriyev (Central Bank of the Republic of Azerbaijan); Aygun Garayeva (Central Bank of the Republic of Azerbaijan); Gulzar Tahirova (Central Bank of the Republic of Azerbaijan) |
| Abstract: | Using 800, 000 transaction-level customs records from January 2018 to February 2026, the paper constructs a trade-weighted Imported Food Price Index (IFPI), covering 34 items from the consumer basket with significant import dependence. The index is developed using the Fisher ideal methodology to provide a timely measure of external food price pressures. The results indicate that the IFPI leads official food Consumer Price Index (CPI) by approximately two months, with a maximum correlation of 0.81, highlighting its potential usefulness as an early indicator of domestic food inflation. Building on this, the paper develops a forecasting framework for the IFPI by combining non-parametric Binary Segmentation and Hidden Markov Models with a regularized machine-learning ensemble. The model employs an ensemble approach that combines Histogram-based Gradient Boosting Regression Tree, Random Forest, and Extreme Gradient Boosting, alongside rigorous time-series crossvalidation. The optimized ensemble achieves a 58% out-of-sample R² relative to a random walk benchmark, vastly outperforming traditional linear Autoregressive Distributed Lag (ARDL) (13.60%) and Autoregressive Integrated Moving Average with Exogenous Variables (ARIMAX) (0.18%) baselines. The forecast results are intended to be incorporated into broader inflation forecasting models to improve short-term projections. |
| Keywords: | Import price index; Fisher Ideal index; Food price inflation; Machine learning forecasting; Hidden Markov models |
| JEL: | C43 C53 C55 E31 F14 |
| Date: | 2026–08–03 |
| URL: | https://d.repec.org/n?u=RePEc:gii:giihei:heidwp20-2026 |