Financial sentiment analysis enhances market understanding. However, standard Natural Language Processing (NLP) approaches encounter significant challenges when applied to small datasets. This study presents a comparative evaluation of embedding-based techniques for financial news sentiment classification in resource-constrained environments. Word2Vec, GloVe, and sentence transformer representations are evaluated in combination with gradient boosting on a manually labeled dataset of 349 financial news headlines. Experimental results identify a substantial gap between validation and test performance. Despite strong validation metrics, models underperform relative to trivial baselines. The analysis indicates that pretrained embeddings yield diminishing returns below a critical data sufficiency threshold. Small validation sets contribute to overfitting during model selection. Practical application is illustrated through weekly sentiment aggregation and narrative summarization for market monitoring. Overall, the findings indicate that embedding quality alone cannot address fundamental data scarcity in sentiment classification. Practitioners with limited labeled data should consider alternative strategies, including few-shot learning, data augmentation, or lexicon-enhanced hybrid methods.
Figures & tables
Column
Description
Date
Trading date associated with the news headline
News
Text of the headline or news snippet
Open
Opening price of the asset
High
Highest price of the day
Low
Lowest price of the day
Close
Adjusted closing price
TABLE I : Dataset Column Descriptions
Size
Value
Records
349
Sampling
One record per trading day
Text length
Short or medium
Coverage
Continuous date range
Missing Data
Minimal
Labels
Provided for supervised learning
TABLE II : Dataset Size and Properties
Model
Validation
Test
Accuracy
Precision
Recall
F1 Score
Accuracy
Precision
Recall
F1 Score
Word2Vec – Base
0.381
0.381
0.381
0.371
0.238
0.196
0.238
0.213
Word2Vec – Tuned
0.381
0.269
0.381
0.315
0.310
0.235
0.310
0.250
GloVe – Base
0.476
0.406
0.476
0.438
0.452
0.458
0.452
0.441
GloVe – Tuned
0.714
0.758
0.714
0.694
0.429
0.392
0.429
0.388
Sentence Transformer – Base
0.429
0.303
0.429
0.355
0.452
0.347
0.452
0.376
TABLE III : Validation and Test Performance Across Embedding Methods
Fig. 1 : Confusion matrices for tuned embedding models on validation and test sets, showing strong positive-class bias and limited discrimination of neutral and negative sentiment.
Week of January 6, 2019
Positive Events:
∙ Roku Inc announced plans to offer premium video channels…
∙ FDIC Chair expressed no concern over market volatility…
∙ Chinese central bank reduced reserve ratio, freeing 116.5B yuan…
Negative Events:
∙ Apple cut quarterly revenue forecast for first time in 15+ years…