Here, the majority of deals (99%) are classified as "hold," while the remaining 1% comprises both short and long positions. However, for the purpose of this competition, the distinction between short and long is not significant. What truly matters is identifying transactions that are not classified as "hold". These non-hold transactions are represented as a single class labeled as "1".
Dataset: The dataset provided is carefully curated, encompassing a range of features and capturing the dynamics of the stock market. With the majority of transactions falling under the "hold" class, the dataset's imbalance poses a significant hurdle in accurately detecting the non-hold anomalies. Exploring and analyzing the dataset thoroughly will be crucial in gaining insights and devising effective strategies for anomaly detection.
Objective: Your ultimate goal in this competition is to build innovative models that can accurately identify non-hold anomalies within the stock market dataset. Submissions will be evaluated based on key anomaly detection metrics that account for the class imbalance, such as precision, recall, and F1-score.