NIFTY Financial News Headlines Dataset
NIFTY Financial News Headlines Dataset
The NIFTY Financial News Headlines Dataset is a curated resource for financial market forecasting using large language models (LLMs). It comprises two versions: NIFTY-LM for supervised fine-tuning with stock movement labels (Rise, Fall, Neutral) and NIFTY-RL for preference-based alignment (e.g., RLHF). The dataset aggregates news headlines from 2010 to 2020, filtered via topic modeling (SBERT) and TF-IDF, and includes market context (OHLCV, technical indicators) for each day. With 2,111 samples, it provides a high-quality benchmark for LLM-based prediction.
Methodologically, the authors demonstrate the dataset's utility through supervised fine-tuning of Llama-2 and Llama-3 models using LoRA adapters, achieving improved accuracy (e.g., 0.45 for Llama-2-7b-chat) compared to other financial datasets. They also analyze LLM embeddings, showing that larger models (by parameter count) and encoder-decoder architectures (T5) yield richer embeddings with higher information gain for tasks like location classification and market movement prediction. The dataset is publicly available on Hugging Face, facilitating research in financial NLP, regime-switching, and LLM alignment.
Highlights
- 1Introduces the NIFTY dataset, a comprehensive financial news headlines dataset spanning 2010-2020, designed for LLM-based market forecasting.
- 2Provides two dataset versions: NIFTY-LM for supervised fine-tuning and NIFTY-RL for preference-based alignment (e.g., RLHF).
- 3Includes extensive metadata, market indices, and deduplicated, filtered news headlines curated via topic modeling and similarity filtering.
- 4Demonstrates that larger LLM embeddings yield richer information gain for financial tasks, with encoder-decoder architectures performing best.
- 5Offers baseline experiments on stock movement prediction using Llama-2 and Llama-3 models, showing improved accuracy after fine-tuning on NIFTY.
Methods
- MData aggregation via web scraping, topic modeling (SBERT similarity), TF-IDF filtering, and deduplication to curate financial news headlines.
- MSupervised fine-tuning (SFT) of LLMs (Llama-2, Llama-3) using LoRA adapters on NIFTY-LM for stock movement classification.
- MPreference dataset construction (NIFTY-RL) with chosen/rejected labels for RLHF-style alignment training.
- MEmbedding analysis using t-SNE and HDBSCAN clustering to measure information gain across model sizes and architectures (GPT-2, BERT, T5).
Results
- RNIFTY-LM contains 2,111 samples (2010-2020) with Rise/Fall/Neutral labels; NIFTY-RL provides preference pairs for alignment.
- RFine-tuning Llama-2-7b-chat on NIFTY improves accuracy from 0.26 to 0.45 on the test split, surpassing other financial datasets.
- RLlama-3-8B-Instruct fine-tuned on NIFTY achieves 0.40 accuracy and 0.30 F1 score, outperforming base model and other datasets.
- RLarger models (GPT-2 Large, BERT Large, T5 Large) produce embeddings with higher information gain for location and market movement tasks.
- REncoder-decoder architectures (T5) yield richer embeddings than encoder-only (BERT) or decoder-only (GPT-2) models of similar size.
Analyze Paper
Generate insights from "NIFTY Financial News Headlines Dataset".