Researchers have developed PRICE, a systematic approach for adapting Large Language Models (LLMs) to forecast Bitcoin prices. This method, applied to a 4-bit quantized LLaMA-3 8B model, integrates parameter-efficient fine-tuning with LoRA, recursive multi-step inference, integer-rounded numerical representation, context-task-format prompting, and exact zero-temperature decoding. Ablation studies indicate that each component enhances forecasting accuracy and reliability, with CTF prompting outperforming Chain-of-Thought and zero-temperature decoding improving stability. PRICE achieved superior performance compared to eight transformer-based and time-series foundation models, demonstrating the critical role of adaptation choices in LLM numerical time-series forecasting. AI
IMPACT Demonstrates that careful LLM adaptation techniques can yield competitive forecasting performance even for models not primarily trained on time-series data.
RANK_REASON Academic paper detailing a new methodology for LLM adaptation in a specific forecasting task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
