Not All News Is Equal: Topic- and Event-Conditional Sentiment from Finetuned LLMs for Aluminum Price Forecasting
2026
Read Not All News Is Equal: Topic- and Event-Conditional Sentiment from Finetuned LLMs for Aluminum Price Forecasting (2026) 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-2603-09085.txt.
Introduction
Alvaro Paredes Amorin and colleagues ask when news tone helps foresee aluminum prices, building monthly scores from English and Chinese headlines with a tuned Qwen3 model. They mix those scores with metal indices, exchange rates, inflation, and energy prices in long-short tests on a Shanghai exchange. In wild, high-swing periods the tone-fed model clearly beats plain data-only rivals. The gain also shifts with news source, topic, and event kind.
Catalog record
| YEAR | 2026 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 72478 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.09085
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Not All News Is Equal: Topic- and Event-Conditional Sentiment from Finetuned LLMs for Aluminum Price Forecasting Abstract By capturing the prevailing sentiment and market mood, textual data has become increasingly vital for forecasting commodity prices, particularly in metal markets. However, the effectiveness of lightweight, finetuned large language models (LLMs) in extracting predictive signals for aluminum prices—and the specific market conditions under which these signals are most informative—remains under-explored. This study generates monthly sentiment scores from English and Chinese news headlines (Reuters, Dow Jones Newswires, and China News Service) and integrates them with traditional tabular data, including base metal indices, exchange rates, inflation rates, and energy prices. We evaluate the predictive performance and economic utility of these models through long-short simulations on the Shanghai Metal Exchange from 2007 to 2024. Our results demonstrate that during periods of high volatility, Long Short-Term Memory (LSTM) models incorporating sentiment data from a finetuned Qwen3 model (Sharpe ratio 1.04) significantly outperform baseline models using tabular data alone (Sharpe ratio 0.23). Subsequent analysis elucidates the nuanced roles of news sources, topics, and event types in aluminum price forecasting Keywords: aluminum price, natural language processing, large language model, sentiment analysis, finance Not All News Is Equal: Topic- and Event-Conditional Sentiment from Finetuned LLMs for Aluminum Price Forecasting Alvaro Paredes Amorin 1,2 , Andre Python 2,3,4 , Christoph Weisser 5 1 International Business School, Zhejiang University 2 Center for Data Science, Zhejiang University 3 Centre for Human Genetics, Nuffield Department of Medicine, Oxford University 4 School of Medicine, Zhejiang University 5 Bielefeld School of Business, Hochschule Bielefeld (HSBI) - University of Applied Sciences and Arts [email protected] [email protected] [email protected] Abstract content 1. Introduction Aluminum is a key non-ferrous metal with widespread applications in automotive, aerospace, construction, and electronics as a result of its lightweight, high corrosion resistance, and excellent conductivity. Aluminum production represents approximately 3.5% of the electricity consumed worldwide and contributes to approximately 1% of global carbon emissions, making it a highly energy-intensive and strategically important commodity ( 6 ; 37 ; 18 ) . Its price dynamics is influenced by a combination of supply constraints, energy costs, geopolitical developments, and demand from emerging industries such as electric vehicles and renewable energy infrastructure ( 27 ; 1 ) . These factors contribute to an increase in price volatility, which presents challenges for both market participants and industrial decision-makers. Aluminum price forecasting typically uses statistical and machine learning methods that relies on tabular (numerical) data on historical prices within a time series modeling framework ( 33 ; 9 ; 29 ) . Although methods that exclusively use tabular data can capture some patterns in price variations, they cannot capture information from textual data sources, such as news headlines and analyst reports, which can provide complementary information that reflect, e.g., investor sentiment and expectations. We investigate whether the sentiment derived from finetuned large language models (LLMs) can improve aluminum price prediction and inform trading strategies. Specifically, we construct sentiment variables from English and Chinese news headlines, classify them with finetuned LLMs, and integrate these signals with numerical time-series data to forecast monthly aluminum prices. Our study also assesses how news sources, topics, and event types can lead to variation in the relevance of signals. By combining time-series models with finetuned LLM sentiment, we identify the conditions under which textual information provides the greatest economic value, offering both methodological insights and practical guidance for commodity market forecasting. 2. Related Work 2.1. Statistical and machine learning approaches to forecast metal prices Non-ferrous prices have been commonly forecasted using statistical models applied to time series data, such as Runge Kutta methods 33 and autoregressive integrated moving average (ARIMA) models 8 ; 21 . More recently, 28 showed that tree-based algorithms, such as extreme gradient boosting (XGBoost), can outperform statistical methods in predicting aluminum prices. 29 showed that auto-regressive light gradient-boosting machine models can further improve the predictive performance on aluminum returns over six months. Larger models based on recurrent neural network architectures, such as long-short term memory (LSTM) models 15 —a type of neural network model particularly effective at capturing long-term dependencies within time series data 16 —have shown promising results in predicting aluminum prices 9 . Both models integrating several machine learning techniques, also-called hybrid models 25 , and models aggregating the outputs of multiple models, which refer to ensemble models 9 , showed promising results in the forecast of aluminum and other non-ferrous metal prices. 2.2. Role of textual data in stock price forecast Understanding market sentiment can provide a valuable context for interpreting analyst forecasts 4 . For example, 22 ; 35 ; 19 find that the integration of sentiment scores with traditional financial indicators can improve stock price forecasts. When combined with data on copper and aluminum commodities prices, 32 found that sentiment analysis can help predict the value of metal industry companies on the stock market. While 5 showed that social media sentiment improves the accuracy of traditional econometric models for copper price forecasting, 13 found that sentiments derived from Twitter and news articles positively correlates with short-term prices of gold and silver. Within a hybrid model framework 38 added a module on the opinion score from a Chinese news website that improved the predictive accuracy of short-term gold prices. 2.3. A growing influence of large language models Large language models (LLMs) play an increasing role in the prediction of financial assets. Without the need for explicit rules, LLMs can capture sentiments from various languages by learning contextual representations via pre-training on massive corpora. Several studies demonstrate that finetuned LLM-based sentiment signals outperform traditional financial NLP benchmarks when used in forecasting tasks. 23 show that an LLaMA3 model finetuned on FinancialPhraseBank leads to a superior sentiment classification performance compared to the base LLaMA3 model. 40 further report that instruction-tuned LLaMA 7B models provide more informative signals for downstream financial prediction tasks than non-instructed variants. Similarly, 39 find that combining fine-tuning with retrieval-augmented generation improves the economic relevance of model outputs. Applying this to a portfolio comprised of 417 stocks from the S&P 500, 20 obtained the highest returns with their finetuned LLaMA2 model, outperforming other lexicon-based methods and FinBERT. The effectiveness of adaptation techniques is also observed across learning regimes. 12 show that fine-tuning Flan-T5 models significantly improves forecasting-related performance compared to zero- and few-shot setups, while adapted open-source LLMs such as LLaMA3, Mistral and Phi consistently outperform finance baseline models. Comparable gains from fine-tuning are reported by 36 in multiple open-source LLMs, with LLaMA2 and MPT producing the strongest downstream results. Finetuned LLMs can efficiently capture market-relevant information not only from English but also from, e.g., Chinese textual data ( 24 ) . More recently, 30 show that DeepSeek, Qwen, and LLaMA models fine-tuned on English and Chinese languages can consistently outperform the benchmark financial model FinBERT, which is a pre-trained BERT-based language model specifically designed to analyze sentiment and extract information from financial text. 3. Methods To assess the role of sentiment from news data in predicting closing aluminum prices, we compare several time series forecasting models that use tabular (“Data 1”) and/or sentiment (“Data 2”) data. The general workflow is summarized in Figure 1 . Figure 1: Workflow. Data 1 : financial data from WIND