The Media Effect: Why Where You Get Your News Matters for AI Stock Prediction

Executive Summary

TL;DR: In the world of FinTech, not all news is created equal. This research demonstrates that a Deep Learning model trained on specialized financial news (MoneyDJ) can achieve a 22.43% ROI, vastly outperforming general news sources like AppleDaily. By combining a custom-expanded Chinese financial lexicon (iMFinanceSD) with Deep Neural Networks, the authors prove that domain-specific media expertise is a critical feature for predictive accuracy.

Context: This paper sits at the intersection of Natural Language Processing (NLP) and Behavioral Finance. It marks a transition from traditional dictionary-based sentiment analysis to the era of Deep Learning (DL), emphasizing that the source of information is as important as the content.

The Problem: The "Noise" in General Media

Investors are bombarded with hundreds of news articles daily. However, previous studies often treated all news sources as a monolithic block. The authors identify a significant gap:

1. Media Bias: Different outlets have different editorial standards and industry knowledge.
2. Lexical Scarcity: General-purpose Chinese sentiment dictionaries (like NTUSD) fail to capture the nuances of financial jargon where words like "shock" or "inject" have specific market meanings.

Methodology: Beyond Simple Word Counting

The research utilizes a sophisticated pipeline that bridges the gap between traditional linguistics and modern AI.

1. The iMFinanceSD Lexicon

To solve the vocabulary problem, the authors integrated existing lexicons and used a Suffix Array algorithm to detect high-frequency "unknown" words in their corpus. This resulted in the iMFinanceSD, a specialized dictionary containing terms like "stabbed sharply" (sharp price drop) and "stabilize."

2. Deep Learning Architecture

Instead of simple linear regression, the study employs a Deep Neural Network (DNN).

• Features: 11 key eigenvalues, including positive/negative word counts and word differences across multiple lexicons.
• Activation: The use of ReLU (Rectified Linear Unit) was pivotal here for speeding up training and handling non-linear relationships between sentiment and market reaction.

Figure 4: The Deep Neural Network structure used to map sentiment features to market trends.

Experimental Results: The MoneyDJ Superiority

The experiment covered 8,472 articles across 18 public companies over two years. The results were categorized by prediction windows (5, 20, and 60 days).

• The Winner: MoneyDJ, a dedicated financial news provider.
• The Quantitative Leap: Using the Deep Learning model on MoneyDJ data for a 60-day strategy yielded a 22.43% ROI.
• General vs. Specific: In contrast, general media like AppleDaily only reached a max ROI of around 1.5%, which the authors deem "unqualified for investment strategies."

Figure 17: ROI Heatmap showing the clear dominance of specialized news providers in longer-term (60-day) forecasts.

Critical Insight & Conclusion

The core takeaway is that Information Efficiency varies by source. Financial news providers like MoneyDJ act as a filter, providing higher signal-to-noise ratios that Deep Learning models can exploit effectively.

Limitations: The study relies on manual feature engineering (11 eigenvalues). In the modern era of LLMs, we would expect "End-to-End" learning where the model reads the raw text directly. However, the fundamental insight remains: Source-weighting is essential. For developers building trading bots today, the lesson is clear—prioritize your data scrapers for specialized domains rather than casting a wide, shallow net.