|
on Big Data |
| By: | Matthias Soot; Daniel Kretzschmar; Alexandra Weitkamp |
| Abstract: | The year of construction is one of the most important building features, indicating factors such as building morphology, building material types and construction technologies. Knowing the specific year of construction of a building helps to estimate energy consumption and demolition waste, answers questions about urban renewal processes in the context of urban planning, improves the accuracy of material stock models and material cadasters and advances MFA and LCA. Additionally, the building age is relevant for real estate value and rent estimations. In Germany, detailed information for individual buildings is not publicly available due to data protection regulations. Traditionally, this information is collected by experts who roughly classify buildings into an age category based on specific visual building characteristics. In recent years, the possibilities of extracting construction age information from the visual appearance of building façades via ground image data have been increasingly explored via deep convolutional neural networks (DCNN). An easy way to derive Information about images are pretrained Visual Language Models (VLM) and Large Language Models (LLM). The question arises as to whether the laborious estimation of the age of a building by humans still guarantees better results in times of freely accessible LLMs, or whether AI can beat the experts. In our study, we conducted a comprehensive survey of more than 350 real estate experts, each of whom was asked to evaluate five randomized images. Analogously, we asked the AI to do the same by using image recognition to ask ChatGPT to estimate the age of construction. The results show that the AI achieves significantly better results than individual experts across all building age groups. Only by sorting the experts according to their experience, local expertise and their own expertise assessment, and by combining the experts' estimations, certain groups of experts manage to beat the AI. |
| Keywords: | building age; construction year; large language models; real estate experts |
| JEL: | R3 |
| Date: | 2025–01–01 |
| URL: | https://d.repec.org/n?u=RePEc:arz:wpaper:eres2025_262 |
| By: | Lampe, Max; Adalid, Ramón |
| Abstract: | Monetary aggregates provide valuable information about the monetary policy transmission and the business cycle. This paper applies machine learning methods, namely Learning Vector Quantisation (LVQ) and its distinction-sensitive extension (DSLVQ), to identify turning points in euro area M1 and M3. We benchmark performance against the Bry–Boschan algorithm and standard classifiers. Our results show that LVQ detects M1 turning points with only a three-month delay, halving the six-month confirmation lag of Bry–Boschan dating. DSLVQ delivers comparable accuracy while offering interpretability: it assigns weights to the sources of broad money growth, showing that lending to households and firms, as well as Eurosystem asset purchases when present, are the main drivers of turning points in M3. The findings are robust across parameter choices, bootstrap designs, alternative performance metrics, and comparator models. These results demonstrate that machine learning can yield more timely and interpretable signals from monetary aggregates. For policymakers, this approach enhances the information set available for assessing near-term economic dynamics and understanding the transmission of monetary policy. JEL Classification: E32, E51, C63 |
| Keywords: | machine learning, monetary aggregates, turning points |
| Date: | 2025–11 |
| URL: | https://d.repec.org/n?u=RePEc:ecb:ecbwps:20253148 |
| By: | Nikolas Anic (Swiss Finance Institute - University of Zurich; Finreon); Andrea Barbon (University of St. Gallen; University of St.Gallen); Ralf Seiz (University of St.Gallen; Finreon); Carlo Zarattini (Concretum Group) |
| Abstract: | This paper investigates whether large language models (LLMs) can improve cross-sectional momentum strategies by extracting predictive signals from firm-specific news. We combine daily U.S. equity returns for S&P 500 constituents with high-frequency news data and use prompt-engineered queries to ChatGPT that inform the model when a stock is about to enter a momentum portfolio. The LLM evaluates whether recent news supports a continuation of past returns, producing scores that condition both stock selection and portfolio weights. An LLM-enhanced momentum strategy outperforms a standard longonly momentum benchmark, delivering higher Sharpe and Sortino ratios both in-sample and in a truly out-of-sample period after the model's pre-training cutoff. These gains are robust to transaction costs, prompt design, and portfolio constraints, and are strongest for concentrated, high-conviction portfolios. The results suggest that LLMs can serve as effective real-time interpreters of financial news, adding incremental value to established factor-based investment strategies. |
| Keywords: | Large Language Models, Momentum Investing, Textual Analysis, News Sentiment, Artificial Intelligence |
| Date: | 2025–10 |
| URL: | https://d.repec.org/n?u=RePEc:chf:rpseri:rp2594 |
| By: | Cansu Isler |
| Abstract: | Accurate forecasting of downside risks to economic growth is critically important for policymakers and financial institutions, particularly in the wake of recent economic crises. This paper extends the Growth-at-Risk (GaR) approach by introducing a novel daily sentiment indicator derived from textual analysis of mandatory corporate disclosures (SEC 10-K and 10-Q reports) to forecast downside risks to economic growth. Using the Loughran--McDonald dictionary and a word-count methodology, I compute firm-level tone growth as the year-over-year difference between positive and negative sentiment expressed in corporate filings. These firm-specific sentiment metrics are aggregated into a weekly tone index, weighted by firms' market capitalizations to capture broader, economy-wide sentiment dynamics. Integrated into a mixed-data sampling (MIDAS) quantile regression framework, this sentiment-based indicator enhances the prediction of GDP growth downturns, outperforming traditional financial market indicators such as the National Financial Conditions Index (NFCI). The findings underscore corporate textual data as a powerful and timely resource for macroeconomic risk assessment and informed policymaking. |
| Date: | 2025–11 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2511.04935 |
| By: | Toby Barter; Zheng Gao; Eva Christodoulaki; Jing Chen; John Cartlidge |
| Abstract: | Bond markets respond differently to macroeconomic news compared to equity markets, yet most sentiment models, including FinBERT, are trained primarily on general financial or equity news data. This mismatch is important because bond prices often move in the opposite direction to economic optimism, making general or equity-based sentiment tools potentially misleading. In this paper, we introduce BondBERT, a transformer-based language model fine-tuned on bond-specific news. BondBERT can act as the perception and reasoning component of a financial decision-support agent, providing sentiment signals that integrate with forecasting models. It is a generalisable framework for adapting transformers to low-volatility, domain-inverse sentiment tasks by compiling and cleaning 30, 000 UK bond market articles (2018--2025) for training, validation, and testing. We compare BondBERT's sentiment predictions against FinBERT, FinGPT, and Instruct-FinGPT using event-based correlation, up/down accuracy analyses, and LSTM forecasting across ten UK sovereign bonds. We find that BondBERT consistently produces positive correlations with bond returns, achieves higher alignment and forecasting accuracy than the three baseline models, with lower normalised RMSE and higher information coefficient. These results demonstrate that domain-specific sentiment adaptation better captures fixed income dynamics, bridging a gap between NLP advances and bond market analytics. |
| Date: | 2025–10 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2511.01869 |
| By: | Leonardo N. Ferreira; Caio Garzeri; Diogo Guillen; Antônio Lima; Victor Monteiro |
| Abstract: | This paper quantifies the “prediction value” of different forms of central bank communication. Combining traditional econometrics and natural language processing, we test how much forecast-improving information can be extracted from the different layers of the Federal Reserve communication. We find that committee-wise communication (statements and minutes) and speeches by the Chair and the Vice Chair improve interest rate forecasts, suggesting that they provide additional information to understand the policy reaction function. However, individual communication beyond the Vice Chair, such as speeches by board members, other FOMC members, and Federal Reserve Bank presidents not sitting in FOMC, is not forecast improving and sometimes even worsens interest-rate forecasts. Based on our theoretical model, we interpret these results as suggesting that the Fed may have overcommunicated, providing excessive noise-inducing communication for forecasting purposes. |
| Date: | 2025–11 |
| URL: | https://d.repec.org/n?u=RePEc:bcb:wpaper:635 |
| By: | Felipe Valencia-Clavijo |
| Abstract: | Large language models (LLMs) are increasingly examined as both behavioral subjects and decision systems, yet it remains unclear whether observed cognitive biases reflect surface imitation or deeper probability shifts. Anchoring bias, a classic human judgment bias, offers a critical test case. While prior work shows LLMs exhibit anchoring, most evidence relies on surface-level outputs, leaving internal mechanisms and attributional contributions unexplored. This paper advances the study of anchoring in LLMs through three contributions: (1) a log-probability-based behavioral analysis showing that anchors shift entire output distributions, with controls for training-data contamination; (2) exact Shapley-value attribution over structured prompt fields to quantify anchor influence on model log-probabilities; and (3) a unified Anchoring Bias Sensitivity Score integrating behavioral and attributional evidence across six open-source models. Results reveal robust anchoring effects in Gemma-2B, Phi-2, and Llama-2-7B, with attribution signaling that the anchors influence reweighting. Smaller models such as GPT-2, Falcon-RW-1B, and GPT-Neo-125M show variability, suggesting scale may modulate sensitivity. Attributional effects, however, vary across prompt designs, underscoring fragility in treating LLMs as human substitutes. The findings demonstrate that anchoring bias in LLMs is robust, measurable, and interpretable, while highlighting risks in applied domains. More broadly, the framework bridges behavioral science, LLM safety, and interpretability, offering a reproducible path for evaluating other cognitive biases in LLMs. |
| Date: | 2025–11 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2511.05766 |