Predicting stock market fluctuations is notoriously challenging, even for experienced experts. Despite substantial research efforts to identify the best datasets and methods for forecasting future trends, predicting stock market fluctuations remains daunting. Investors with access to options chain data have a unique advantage in predicting future stock movements. Using a novel approach, this study leverages historical prices and options chain data to predict the stock market's closing price movement. We employ a piece-wise linear encoding technique to convert historical prices and options chain data into a high-dimensional vector embedding. This numerical embedding serves as input for various statistical and neural network classifiers evaluated based on their accuracy and Matthew correlation coefficient (MCC). Furthermore, we introduce a cosine similarity-based strategy that does not require training, unlike conventional machine learning and deep learning models. This approach relies on finding comparable vectors in past data to forecast stock movements accurately. The proposed model's results demonstrate a substantial improvement in both accuracy and MCC scores over the best-performing baseline methods, highlighting the effective utilisation of numerical embedding for precise predictions of closing price movements. Furthermore, the model exhibits consistent performance gains with increasing data volume and significantly outperforms state-of-the-art methods in terms of performance, training and inference speed, making it both scalable and computationally efficient.

The financial markets, particularly stock trading, offer a variety of profit-generating opportunities based on complex and volatile behaviour. Investors seek strategies to maximise returns, leading to an investigation of inherent market patterns. Converting OHLC (Open, High, Low, Close) data into transformers-based pre-trained language model compatible text is an innovative method for representing numeric data. Extending the language model’s utility to integrate stock market numeric time-series data incorporates its inherent numeracy in embeddings. Raw data are converted into a format compatible with the pre-trained language model through preprocessing and text templates. Using an ensemble of Bidirectional Encoder Representations from Transformers (BERT), FinBERT (BERT finetuned with the financial corpus), FLANG-BERT (BERT finetuned with the financial corpus) and FLANG-ELECTRA (ELECTRA finetuned with the financial corpus) as feature extractor, historical stock market data are utilised to generate an embedding matrix and fused with established neural network architectures, such as Backpropagation Neural Network (BPNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), to predict stock market trends. The simulation results demonstrate that the proposed integrated approach is preferable to previous methodologies. The significance of the findings is confirmed by statistical validation using the Wilcoxon signed-rank test (p value < 0.01). This study offers a promising approach for improving stock market trend prediction by integrating the ensemble of language model-based numeric embeddings with neural networks.

Misinformation circulating on social platforms often distorts public understanding and can escalate real-world consequences, particularly in volatile geopolitical contexts. In this study, we describe the system constructed for subtask 3 of the PROMID 2025 shared task, which focuses on identifying misleading content within tweets pertaining to the Russo-Ukrainian conflict. The submission from our team (Sarang) secured 2nd place, supported by strong evaluation metrics: a precision of 0.90, a recall of 0.91, and a weighted F1-score of 0.90. Because the dataset contained instances in multiple languages, all non-English posts were rendered into English through a large language model. To mitigate skewed label distributions, we introduced synthetic variants of the minority class, thereby easing class imbalance. The classification pipeline relied on a deberta-v3-small encoder, which proved adept at capturing fine-grained semantic distinctions. The resulting performance underscores the reliability of the proposed approach and provides a competitive reference point for forthcoming work. Overall, the study offers practical insights for advancing misinformation detection in future shared-task settings.

This paper describes our approach for the FinCausal 2025 English Shared Task, aimed at detecting and extracting causal relationships from the financial text. The task involved answering context-driven questions to identify causes or effects within specified text segments. Our method utilized a consciousAI RoBERTa-base encoder model, fine-tuned on the SQuADx dataset. We further fine-tuned it using the FinCausal 2025 development set. To enhance the quality and contextual relevance of the answers, we passed outputs from the extractive model through Gemma2-9B, a generative large language model, for answer refinement. This hybrid approach effectively addressed the task’s requirements, showcasing the strength of combining extractive and generative models. We (Team name: Sarang) achieved outstanding results, securing 3rd rank with a Semantic Answer Similarity (SAS) score of 96.74% and an Exact Match (EM) score of 70.14%.