Profiting from Chaos: Finding Signal in a Reality
Financial markets are constantly surrounded by information, with thousands of news articles describing economic events and corporate developments daily. The challenge within this overwhelming flow is to find the signal hidden within the noise. In this competition, participants are challenged to develop a machine learning model that can understand and classify the sentiment of financial news articles.
The competition is designed as a multi-class text classification problem. Participants will receive a labeled training dataset containing financial news articles and their corresponding sentiment labels. A separate test dataset will be provided without labels. Your mission is to explore, understand, model, predict, and discover the signal.
Participants are encouraged to explore different representations of the text and investigate how article content, headlines, publication information, and other available features can contribute to sentiment prediction. You are free to experiment with various approaches, including: * Traditional machine learning * Natural Language Processing (NLP) * Text feature engineering * Embedding-based approaches * Deep learning * Pre-trained language models * Ensemble methods
The competition provides a collection of financial and economic news articles: * Training Data: 2,692 labeled articles. * Test Data: 674 unlabeled articles.
Each article contains the following information:
* date_publication: The publication date of the article.
* media_source: The media outlet that published the article.
* title_article: The title or headline of the article.
* content_article: The full content of the article.
* sentiment_label: The sentiment label (available only in training data).
The sentiment labels consist of three classes: POSITIF, NETRAL, and NEGATIF.
The preliminary round consists of two components: 1. 70% — Kaggle Leaderboard Performance: Submissions are evaluated on the hidden test data (divided into a Public Leaderboard of 202 articles and a Private Leaderboard of 472 articles). 2. 30% — Notebook Quality: The submitted Kaggle Notebook is assessed based on data understanding, exploratory analysis, preprocessing, feature engineering, modeling methodology, validation strategy, clarity of explanation, quality of insights, and reproducibility.
A strong solution requires more than just a high leaderboard score; it must clearly communicate what was done, why it was done, and what was learned from the data.