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

This study demonstrates that integrating sentiment scores derived from a finetuned Qwen3 model analyzing English and Chinese news significantly enhances aluminum price forecasting accuracy and economic utility, particularly during periods of high market volatility, compared to traditional tabular data models.

The Gist

Imagine you are trying to guess the price of a barrel of aluminum next month.

Traditionally, financial experts have been like weather forecasters looking only at a thermometer. They stare at numbers: past prices, interest rates, and how much energy costs. They use complex math to see if the "temperature" is going up or down. This works okay, but it's like trying to predict a storm by only looking at the thermometer and ignoring the dark clouds gathering on the horizon.

This paper asks a simple question: What if we also read the news?

The authors decided to build a "super-reader" using a type of Artificial Intelligence called a Large Language Model (specifically, a tuned version of Qwen3). This AI doesn't just read numbers; it reads headlines from Reuters, Dow Jones, and Chinese news sources to understand the mood of the market. Is everyone panicking? Are they excited? Is there a rumor of a factory shutdown?

Here is the breakdown of their findings, using some everyday analogies:

1. The "Calm Sea" vs. The "Storm"

The researchers found that the value of reading the news depends entirely on how rough the ocean is.

• When the market is calm (Low Volatility): Everyone is just driving on a straight road. Looking at the speedometer (past prices) is enough. Adding news is like trying to steer a car by listening to the radio while driving on a straight highway; it doesn't help much and might even distract you.
• When the market is stormy (High Volatility): This is where the magic happens. When prices are swinging wildly (like during a crisis or a sudden supply shock), the "thermometer" (past data) lags behind. It's like looking at a rearview mirror while driving into a sudden fog.
  • The Result: The AI that reads the news acts like a spotter on the roof of the car. It sees the pothole or the storm coming before the car hits it. In these turbulent times, the model using news sentiment was 3.5 times better at making profitable trades than the model that only looked at numbers.

2. Not All News is Created Equal (The "Gossip" vs. The "Fact")

The team realized that not every headline is useful. They tested different types of news, like sorting through a pile of mail.

• The "Future Gossip" (Predictions): Headlines like "Analysts think prices might go up next year" are often useless. Why? Because the market has already heard this gossip and priced it in. It's like betting on a horse race after everyone else has already placed their bets. These stories didn't help predict the price.
• The "Hard Facts" (What Happened): Headlines like "A factory in China just shut down due to a fire" or "Inventory levels dropped by 10% today" were gold. These are actual events, not guesses. The AI learned to ignore the "maybe" stories and focus on the "did" stories.
  • The Lesson: If you want to predict the future, focus on what actually happened, not what people think might happen.

3. The Source Matters (The "Reliable Friend" vs. The "Chatterbox")

The researchers compared three news sources: Reuters, Dow Jones, and China News Service.

• Reuters was the reliable friend who tells you the truth clearly and quickly. When the AI read Reuters, it made the most money.
• Dow Jones was like a friend who talks a lot about specific companies but misses the big picture. It focused too much on individual company news and not enough on the big price movements.
• China News Service was good at talking about production numbers but sometimes missed the emotional pulse of the global market.

It turns out, it's not just what you read, but who you read. A clear, direct report on price movements is worth more than a long, complicated story about a single company.

4. The "Secret Sauce" (Fine-Tuning)

The AI they used wasn't just a generic chatbot. They took a powerful model and "fine-tuned" it on thousands of financial documents. Think of it like taking a brilliant generalist doctor and giving them a 10-year residency specifically in aluminum market diseases. This specialized training allowed the AI to understand financial nuance much better than standard tools.

The Bottom Line

This paper proves that in the world of aluminum trading (and likely other commodities), numbers tell you where you've been, but news tells you where you're going.

• If the market is boring and stable, stick to the numbers.
• If the market is chaotic and scary, listen to the news.
• But be smart about it: Ignore the rumors and predictions; focus on the hard facts and get your news from the most reliable source.

By combining the "thermometer" (numbers) with the "spotter" (AI-read news), investors can navigate the stormy seas of commodity trading much more safely and profitably.

Technical Summary

Here is a detailed technical summary of the paper "Not All News Is Equal: Topic- and Event-Conditional Sentiment from Finetuned LLMs for Aluminum Price Forecasting."

1. Problem Statement

Forecasting commodity prices, specifically aluminum, is challenging due to the interplay of supply constraints, energy costs, geopolitical factors, and emerging demand (e.g., EVs). Traditional forecasting relies heavily on tabular numerical data (historical prices, exchange rates, inflation, energy prices) using statistical or machine learning models. However, these models often fail to capture market sentiment and investor expectations embedded in textual data (news headlines, analyst reports).

The specific gaps this study addresses are:

• The effectiveness of lightweight, finetuned Large Language Models (LLMs) in extracting predictive signals from financial text compared to traditional NLP models (like FinBERT).
• The lack of understanding regarding which specific market conditions (e.g., volatility regimes) make sentiment data most valuable.
• The variation in predictive power based on news sources, topics, and event types (forward-looking vs. factual).

2. Methodology

Data Collection and Preprocessing

• Numerical Data (Data 1): Daily closing prices and macroeconomic indicators (exchange rates, inflation, copper/zinc/iron prices, oil/gas prices) from the Shanghai Stock Exchange (Wind terminal), aggregated to monthly frequency from March 2007 to April 2024.
• Textual Data (Data 2): News headlines collected from Factiva (2007–2024) from three sources:
  • Reuters (English, N=4,963)
  • Dow Jones Newswires (English, N=11,581)
  • China News Service (Chinese, N=8,970)
• Filtering: A few-shot LLaMA3 8B model was used to filter out irrelevant entries, ensuring only aluminum-specific news remained.

Sentiment Analysis Pipeline

• Models: Two models were compared for sentiment classification:
  1. FinBERT: A standard BERT-based financial sentiment model.
  2. Finetuned Qwen3 (8B): A lightweight LLM finetuned on five diverse financial sentiment datasets (including FinancialPhraseBank, Twitter Sentiment, and Chinese Finance Sentiment).
• Classification: Headlines were classified as Positive (+1), Negative (-1), or Neutral (0).
• Aggregation: Monthly sentiment scores were calculated as a weighted sum of individual headline scores for each month.

Forecasting and Trading Strategy

• Models: Various time-series architectures were tested, including LSTM, BiLSTM, GRU, ConvLSTM, and TFT (Temporal Fusion Transformers).
• Input Configurations: Models were trained on:
  1. Tabular data only (Baseline).
  2. Tabular data + Sentiment scores.
  3. Sentiment scores only.
• Trading Signal Generation:
  • Sentiment Strategy: Buy if monthly sentiment > 0, Sell if < 0.
  • Price-Based Strategy: Buy if predicted price ($P_{t+1}$) > current price ($P_t$).
• Evaluation: Strategies were backtested on the Shanghai Metal Exchange using Cumulative Return and Sharpe Ratio.
• Volatility Regimes: The sample was partitioned into Low, Medium, and High volatility periods based on a 6-month rolling standard deviation of returns.

3. Key Contributions

1. Demonstration of Finetuned LLM Superiority: The study provides empirical evidence that a finetuned Qwen3 model outperforms FinBERT in extracting sentiment signals for commodity forecasting, particularly in non-English (Chinese) contexts and complex financial narratives.
2. Context-Dependent Value of Sentiment: The paper establishes that the economic utility of sentiment is non-linear and highly dependent on market volatility. Sentiment is not always "better"; its value shifts based on market conditions.
3. Granular Signal Analysis: The research moves beyond "aggregate sentiment" to identify that:
   • Source Quality Matters: Reuters provides higher signal-to-noise ratios than Dow Jones or China News Service.
   • Topic Filtering is Critical: Aggregating all news is suboptimal. Specific combinations of topics (e.g., Company News + Supply Disruption + Price Movement) yield significantly higher Sharpe ratios than a "all-topics" approach.
   • Event Type Distinction: Factual, past-event reporting (e.g., inventory releases, production figures) is far more predictive than forward-looking commentary (e.g., analyst forecasts), which aligns with the Efficient Market Hypothesis.

4. Key Results

Performance by Volatility Regime

• High Volatility: The Tabular + Sentiment strategy using the finetuned Qwen model achieved a Sharpe Ratio of 1.04, a massive improvement over the Tabular-only baseline (0.23). In turbulent markets, sentiment captures panic/recovery signals that numerical models miss.
• Medium Volatility: The Sentiment-Only strategy dominated (Sharpe Ratio 1.19), outperforming both the combined approach (0.51) and the baseline. This suggests that in normal markets, sentiment alone is a sufficient predictor.
• Low Volatility: All strategies converged to similar performance (Sharpe 0.29–0.54), indicating limited differentiation in calm markets.

Source and Topic Analysis

• Source: Reuters headlines yielded the highest Sharpe ratio (0.80) and cumulative returns (433%). Dow Jones (0.18) and China News Service (0.15) performed significantly worse, largely due to topic allocation (e.g., Dow Jones focused heavily on "Company News," which had negative predictive value).
• Topic Optimization: An optimal subset of 8 topics (excluding "Production Output," "Market Analysis," "Trade Policy," and "Environmental") improved the Sharpe ratio by 23.6% compared to using all topics.
• Event Type: Strategies based on past events achieved a Sharpe of 0.62, while forward-looking strategies yielded nearly zero (-0.01).

Model Comparison

• Qwen3 vs. FinBERT: The finetuned Qwen3 consistently generated superior trading signals.
• LSTM Performance: The LSTM model with a 1-month window and Reuters sentiment achieved the highest total return (292%) and a Sharpe ratio of 1.04 in high volatility, significantly outperforming the no-sentiment baseline (131% return).
• Accuracy vs. Profitability: Interestingly, the sentiment-augmented model had slightly worse point-prediction metrics (R², RMSE) than the no-sentiment model but generated much higher trading returns. This indicates that sentiment helps the model correctly predict direction at critical turning points, even if the magnitude of the price prediction is slightly less precise.

5. Significance and Implications

• Methodological Shift: The study validates that finetuned lightweight LLMs are a viable, cost-effective alternative to massive proprietary models for financial NLP, especially when handling multilingual data (English/Chinese).
• Practical Trading Strategy: It offers a concrete framework for commodity traders:
  • Filter by Source: Prioritize high-quality sources like Reuters.
  • Filter by Topic: Exclude noise-heavy topics (e.g., general environmental news) and focus on price dynamics and supply disruptions.
  • Filter by Event Type: Prioritize factual reporting over speculative forecasts.
  • Dynamic Allocation: Weight sentiment inputs heavily during high-volatility periods and rely on sentiment-only signals during medium-volatility periods.
• Economic Insight: The findings challenge the assumption that "more data is better." Instead, curated, high-signal textual data combined with numerical time-series data provides the most robust forecasting capability, particularly for commodities like aluminum which have lower media coverage than gold or silver.

In conclusion, the paper demonstrates that sentiment is not a monolithic signal; its predictive power is conditional on the source, the topic, the event type, and the prevailing market volatility. Integrating these nuanced signals via finetuned LLMs significantly enhances the economic utility of aluminum price forecasting.