Not All News Is Equal: Topic- and Event-Conditional Sentiment from Finetuned LLMs for Aluminum Price Forecasting

Alvaro Paredes Amorin; Andre Python; Christoph Weisser

cs.LGMar 10, 2026

Not All News Is Equal: Topic- and Event-Conditional Sentiment from Finetuned LLMs for Aluminum Price Forecasting

Authors: Alvaro Paredes Amorin, Andre Python, Christoph Weisser

Organizations: International Business School, Zhejiang University · Center for Data Science, Zhejiang University · Centre for Human Genetics, Nuffield Department of Medicine, Oxford University · School of Medicine, Zhejiang University · Bielefeld School of Business, Hochschule Bielefeld (HSBI) - University of Applied Sciences and Arts

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.

Figures & tables

Figure 1: Workflow. Data 1 : financial data from WIND terminal includes tabular data extracted from March 2007 to April 2024 (4,152 rows). Data 2 : textual data that includes headlines from two news sources in English (Reuters (N=4,963), Dow Jones Newswires (N=11,581), and a news source in Chinese (China News Service (N=8,970)) collected from March 2007 to April 2024. The data processing and sentiment analysis (green) includes data scaling, normalization, and the treatment of missing values for the tabular data, and the use of language models to generate new sentiment variables from Data 2 . The sentiment is classified in “positive”, “negative” or “neutral”. Monthly sentiment scores are combined with other numerical data (yellow box) to train and test time series models in order to predict monthly aluminum prices.

Figure 2: Evaluation of portfolio’s performance by strategy and volatility scenarios. For each strategy (tabular-only, tabular+sentiment, sentiment only (qwen), sentiment only (reuters)) the portfolio’s performance is represented by the Sharpe ratio (y-axis) across three volatility scenarios (panels), with: A high volatility scenario (n=28 months), B medium volatility (n=106 months), and C low volatility (n=66 months). Error bars represent ±1 standard error. The highlighted bars (orange) indicate the best performing strategy within each scenario.

Figure 3: Evaluation of topic and event type in predictive aluminum prices. A Comparison of Sharpe ratio across the top five most covered topics (Price Movement, Company News, Production Output, Inventory Stocks, Supply Disruption) and the best topic combination, all calculated using Reuters headlines with the finetuned Qwen sentiment. The benchmark (gray) represents all headlines aggregated. Here $n$ shows the number of months in which each topic is present. B Comparison of forward-looking versus past event news types. C Comparison of the top-performing topic (Price Movement) by event type. Error bars represent ±1 standard error. The green bar indicates the best-performing topic combination, while the orange bars indicate the best performing individual topic or event type.

Figure 4: Predictive performance by topic coverage. A Percentage of headlines allocated to each topic by source. B Global Sharpe ratio by topic, with gold rings highlighting the top three performers. Topics are sorted by global Sharpe ratio (descending).

Topic	Reuters	DowJones	ChinaService	Improvement vs Reuters
	(Benchmark)	Sharpe	Sharpe	DowJones (%)	ChinaService (%)
Price Movement	0.714	0.525	0.674	-26.6	-5.6
Environmental	0.601	0.347	0.395	-42.3	-34.3
Other	0.612	0.123	0.167	-79.9	-72.8
Market Analysis	0.221	0.348	0.020	+57.5	-91.1
Production Output	0.042	0.125	0.283	+198.5	+577.4

Table 1: Sharpe ratio comparison across all individual topics, sorted by global performance. Improvements are computed relative to Reuters as the benchmark.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Cluster plot of the Reuters and Dow Jones Newswires aluminum datasets before (left) and after (right) filtering using LLaMA3. The y and x axis are the 2 dimensions obtained by reducing the embeddings dimensions. The plots on the left show less number of outliers outside the main groups clusters.

Figure 6: Combination of models, sentiment sources and time windows

Figure 7: Average $R^{2}$ score by number of layers and by hidden size. The solid line in black is the average of the five time series models. The faded lines are each respective model $R^{2}$ , blue for BiLSTM, green for TFT, purple for GRU, orange for ConvLSTM and red for LSTM.

Figure 8: Predicted aluminum price and true aluminum price from November 2007 to April 2024.

Figure 9: Sentiment source averaged R 2 score per time window and model on the left, bands show the 95% confidence intervals of those mean R² values. Time window averaged R 2 score per sentiment source and model on the right, whiskers show the 95% confidence intervals of those mean R² values.

A. Model Performance
Configuration		Time Window
Category	Name	1M	3M	6M	12M	Avg
Models	BiLSTM	0.549	0.573	0.562	0.534	0.554
	ConvLSTM	0.585	0.569	0.559	0.543	0.564
	GRU	0.550	0.545	0.562	0.545	0.551
	LSTM	0.564	0.563	0.549	0.548	0.556

Appendix

Table 2: Directional Hit Rate Performance by Model and Sentiment Source Across Time Windows. Hit rates represent the percentage of correct directional predictions. Values closer to 1.000 indicate bet- ter performance. The 3-month window shows the highest average hit rate across both models and sources. Values in bold highlight the time window (panels A and B) with the best average hit rate, and the model (panel A) and sentiment source (B) with the best average hit rate.

A. Model Performance
Configuration		Time Window
Category	Name	1M	3M	6M	12M	Avg	Best
Models	BiLSTM	0.346	0.523	0.097	-0.076	0.222	2.32
	ConvLSTM	-0.292	0.062	-0.096	0.147	-0.045	1.02
	GRU	0.125	-0.308	1.087	-0.334	0.142	2.66
	LSTM	1.687	-0.178	0.105	-0.269	0.336	2.92

Appendix

Table 3: Total return performance by model and sentiment source across time windows. A. Sentiment-source averaged returns for each mode. B. Model-averaged returns for each sentiment source. Values represent total returns (multiplicative factors, where a value of X indicates an X

\times

100% return). The “Avg” column shows time-window averaged performance, and “Best” reports the highest return achieved.

Figure 10: Model averaged returns by time window (A). These are the total returns of all setups averaged for each model and time window. Source averaged returns (B) by time window. These are the total returns of all setups averaged for each sentiment source and time window.

Figure 11: Comparison of the best Sentiment vs No-Sentiment Models returns vs True Aluminum price. Y axis is the Portfolio Value, with the starting point being 100 (100% of the intial value).

Figure 12: Monthly return of best sentiment and no-sentiment models monthly returns. The red bars show month where the two configurations made a different decision.

News Headline	Date	Sentiment
Coronavirus will negatively affect aluminum market in China	28-Feb-20	Negative
Coronavirus double shock for aluminium sector	24-Feb-20	Negative
LME aluminium can test support at $1,688	24-Feb-20	Negative
Japan aluminium stocks down 2.8%	18-Feb-20	Negative
LME aluminium testing support at $1,709	10-Feb-20	Negative
LME aluminium seeking support at $1,709	03-Feb-20	Negative

Appendix

Table 4: Sentiment analysis of market news (February 2020).

News Headline	Date	Sentiment
Norway’s Hydro to cut Slovakia aluminium output further due to power prices	30-Dec-21	Negative
Copper slips in range-bound trade, aluminium shines on supply worries	30-Dec-21	Negative
Power price surge pushes aluminium to 2-month high	23-Dec-21	Positive
China Nov aluminium output at 3.10 mln tonnes – stats bureau	15-Dec-21	Neutral
Scarce supplies to propel aluminium to top LME leaderboard	15-Dec-21	Positive
Marubeni sees Japanese aluminium premiums at $140–$ 250/T in 2022	07-Dec-21	Neutral

Appendix

Table 5: Sentiment analysis of market news (December 2021).

News Headline	Date	Sentiment
London aluminium poised for worst month in over a decade on growth risks	29-Apr-22	Negative
China Shenhuo to raise aluminium output in Yunnan as power curbs ease	26-Apr-22	Positive
Global aluminium output falls 1.55% in March year on year, IAI says	20-Apr-22	Negative
Shanghai aluminium hits 3-month low on strong dollar, demand woes	12-Apr-22	Negative
Shanghai aluminium sinks to 3-month low as demand woes linger	12-Apr-22	Negative
China demand angst hits aluminium prices	11-Apr-22	Negative

Appendix

Table 6: Sentiment analysis of market news (April 2022).

Figure 13: Panel A: Volatility Regimes Over Time. Red area represents the high-volatility regime, yellow represents medium-volatility and green low-volatility. Panel B: Portfolio Values Over Time for the different strategies. Blue line represents the price-based strategy, red line price-based strategy combined with sentiment, green line sentiment only strategy with sentiment predicted by finetuned Qwen, yellow line sentiment only strategy with sentiment predicted by FinBERT and black line buy & hold strategy.

Model	Source	Horizon	Hidden Size	Layers	R 2	RMSE	MAE
bilstm	reuters	1 months	256	5	0.89	126.79	91.59
bilstm	reuters	3 months	256	3	0.88	132.66	92.02
bilstm	reuters	6 months	64	3	0.89	126.92	88.74
bilstm	reuters	12 months	64	3	0.89	128.64	95.47
bilstm	dowjones	1 months	256	5	0.88	130.12	93.37
bilstm	dowjones	3 months	128	4	0.88	129.93	92.71

Appendix

Table 7: Best results (

R^{2}

, RMSE and MAE) for the best hyperparameter combination for each model, sentiment source and time window.

Model	Source	Horizon	R 2	RMSE	MAE	HitRate	p_value	TotRet
lstm	chinaservice	1_months	0.90	121.83	87.18	0.56	0.0665	1.16
lstm	chinaservice	3_months	0.87	138.20	100.94	0.56	0.0651	-0.60
lstm	chinaservice	6_months	0.87	134.75	99.10	0.55	0.1760	-0.08
lstm	chinaservice	12_months	0.88	134.19	98.15	0.55	0.1283	-0.21
lstm	reuters	1_months	0.89	125.12	87.98	0.55	0.1594	2.92
lstm	reuters	3_months	0.87	135.41	94.14	0.60	0.0041	0.45

Appendix

Table 8: Trading simulation results for the best hyperparameter combination for each model, sentiment source and time window. The p-value here measures the statistical significance of the model’s directional accuracy relative to random chance (50%). The TotRet metric shows the total return of the trading strategy over the backtest period.

Explore similar work

PRICE: A Systematic Study of LLM Adaptation Choices for Bitcoin Price Forecasting
Sep 4, 2026Maryam Fakhari, Mehran SafayaniLLM-Based Time Series Forecasting Financial Time Series
Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan, Iran
Beyond Polarity: Multi-Dimensional LLM Sentiment Signals for WTI Crude Oil Futures Return Prediction
Mar 12, 2026Dehao Dai, Ding Ma, Dou Liu +2Financial Forecasting Financial Sentiment Analysis
University of California San Diego, La Jolla, California, United States · Georgia Institute of Technology, Atlanta, Georgia, United States · New York University, New York, New York, United States +2
A Review of Large Language Models for Stock Price Forecasting from a Hedge-Fund Perspective
Apr 10, 2026Olivia Zhang, Zhilin ZhangQuantitative Finance LLM-Based Time Series Forecasting
Hockaday School, 11600 Welch Rd, Dallas, TX 75229, USA · Lumos Alpha, Dallas, TX 75248, USA

Sep 4, 2026J/K move · Enter open · S save

PRICE: A Systematic Study of LLM Adaptation Choices for Bitcoin Price Forecasting

Maryam Fakhari, Mehran Safayani

Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan, Iran

LLM-Based Time Series Forecasting Financial Time Series LLM Fine-Tuning

Open paper AskPDF arXiv

Cryptocurrency markets exhibit extreme volatility and non-stationary dynamics that challenge conventional forecasting methods. Although Large Language Models (LLMs) have shown promise for time series forecasting, the combined effects of adaptation choices remain largely unexplored in financial settings. This study introduces PRICE, a structured approach for adapting LLMs to short-term Bitcoin price forecasting. Built on a 4-bit quantized LLaMA-3 8B model, PRICE investigates how fine-tuning, numerical representation, prompting, inference, and decoding jointly influence forecasting performance. PRICE integrates Parameter-efficient fine-tuning with Low-Rank Adaptation (LoRA), Recursive multi-step inference, Integer-rounded numerical representation, Context-Task-Format (CTF) prompting, and Exact zero-temperature decoding. Ablation studies show that each component contributes to forecasting accuracy and reliability. LoRA enables efficient training on limited hardware, recursive inference improves accuracy, integer-rounded values reduce errors, CTF prompting outperforms Chain-of-Thought, Implicit Chain-of-Thought (iCoT), and few-shot prompting, and zero-temperature decoding improves stability during recursive forecasting. Comparative evaluation against eight transformer-based and time-series foundation models shows that PRICE achieves the lowest forecasting errors on both validation and test sets while maintaining robust performance across evaluation periods. Despite being based on a model primarily pretrained on text rather than time-series data, PRICE achieves competitive or superior performance relative to specialized foundation models. These findings demonstrate that adaptation choices critically determine the accuracy and robustness of LLMs for numerical time-series forecasting.

Abstract

Explore similar work

PRICE: A Systematic Study of LLM Adaptation Choices for Bitcoin Price Forecasting

Beyond Polarity: Multi-Dimensional LLM Sentiment Signals for WTI Crude Oil Futures Return Prediction

A Review of Large Language Models for Stock Price Forecasting from a Hedge-Fund Perspective