Evaluating Financial LLM Architectures on Consumer GPUs: Model Comparison and Temperature-Based Hyperparameter Tuning
Year: 2025
Type: dissertation
Abstract: This thesis investigates whether quantized, domain-specific Large Language Models (LLMs) deployed on consumer-grade hardware can generate profitable trading signals through technical analysis of cryptocurrency markets. We evaluate three models (Finance-LLM-13B, Finance-Llama-8B, and Phi-4) across 45 backtests spanning 15 sessions with heterogeneous hyperparameter configurations on Bitcoin/USDT 5-minute candles from November 2025. Finance-LLM-13B consistently achieves profitability against a -9.04% market baseline. We conduct controlled experiments isolating the effects of temperature and nucleus sampling (top-p) on trading signal quality. Critically, we demonstrate that domain-specific fine-tuning dominates model parameter scaling. Temperature-driven exploration benefits well-trained models but degrades weak models. Results emphasize that domain-specific training, rather than scale or sophisticated decoding strategies, is the dominant performance driver for financial reasoning. This work provides proof-of-concept evidence that LLM-based trading on consumer hardware is feasible, while establishing necessary conditions for robustness: multi-period validation and live-market validation before production deployment.
Author: Kavdir, Bugra
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7208804510