LLMs outperform traditional models in real-time cryptocurrency forecasting

A team of researchers has benchmarked the predictive power of large language models (LLMs) in the fast-moving cryptocurrency sector. Their study, “Large Language Models for Nowcasting Cryptocurrency Market Conditions,” published in FinTech, explores how advanced LLMs can generate near real-time market forecasts, or “nowcasts,” and how they compare with traditional econometric and neural network models.

The research reveals that while LLMs have emerged as promising tools for short-horizon forecasting, especially in relatively stable markets, they face limitations in highly volatile conditions, underscoring the need for hybrid approaches that combine LLM capabilities with established volatility modeling techniques.

Benchmarking LLMs against traditional models

The study evaluated the performance of five leading decoder-only LLMs: GPT-4.1 (OpenAI), Gemini-2.5-Pro (Google), Claude-3-Opus-20240229 (Anthropic), DeepSeek-Reasoner (DeepSeek), and Grok-4 (xAI)  in nowcasting cryptocurrency market conditions. Nowcasting involves using high-frequency data to forecast near-term market movements, a task critical for traders and risk managers dealing with minute-level fluctuations.

To test these models, the researchers analyzed minute-resolution price data for 12 major cryptocurrencies, spanning high-volatility assets such as Bitcoin (BTC) and Ethereum (ETH), mid-cap tokens like Cardano (ADA), Solana (SOL), TRON (TRX), ecosystem tokens such as BNB, STETH, WTRX, and stablecoins including USDT and USDC.

The study compared LLM forecasts against both classical econometric models, including GARCH, GJR-GARCH, and HAR-RV, widely used for volatility modeling, and neural network baselines such as GRU and LSTM, which have been popular in earlier cryptocurrency forecasting efforts.

The findings reveal that Gemini-2.5-Pro consistently outperformed its peers, leading in nowcasting accuracy for 8 out of the 12 assets, particularly excelling in low-volatility environments like stablecoins. LLMs as a group achieved an average accuracy gain of 9.47% over econometric models and 4.63% over neural baselines, showing their advantage in parsing complex market patterns from large-scale data.

Strengths and weaknesses in different market conditions

While LLMs demonstrated strong performance overall, the study highlights that they still face challenges in highly volatile markets, especially for flagship cryptocurrencies such as Bitcoin and Ethereum. These assets often experience abrupt intraday spikes that traditional volatility-focused models are better equipped to handle.

The researchers also tested the models using a multi-pass approach, comparing Pass@1 (single-pass predictions) to Pass@2 (refined multi-pass predictions), which improved stability and reduced prediction errors by an average of 1.22%. This finding points to the importance of iterative reasoning in enhancing LLMs’ performance for time-sensitive financial forecasting.

The study notes that LLMs’ capacity to process vast amounts of unstructured data, such as market sentiment indicators and news trends, is a clear advantage in dynamic market environments. However, the inherent unpredictability of crypto-assets in times of market stress continues to limit their accuracy compared to hybrid methods that incorporate traditional econometric models for volatility forecasting.

Implications for the future of FinTech forecasting

By pioneering the use of LLMs for medium-to-high-frequency nowcasting in cryptocurrency markets, the research opens new possibilities for financial analytics. The authors argue that LLMs, with their advanced reasoning and contextual processing capabilities, have the potential to transform short-horizon forecasting tools used by traders, portfolio managers, and risk analysts.

The study highlights that domain-specific LLMs, models fine-tuned on financial market data, could further enhance predictive accuracy, particularly for handling extreme market volatility and incorporating real-time sentiment data from diverse sources such as social media and global news.

For financial institutions, the findings suggest that integrating hybrid models, combining LLM-driven pattern recognition with the proven strengths of econometric volatility forecasting, could deliver more reliable results, improving both market responsiveness and risk management strategies.

The research also highlights the growing need for robust validation frameworks to ensure that LLM-based forecasts remain transparent, accountable, and resilient to market shocks. As LLMs become increasingly embedded in trading platforms and fintech applications, addressing these governance challenges will be crucial for their adoption in regulated financial environments.