CSTrader Enables Language-Grounded Trading in Niche Asset Markets.

Niche asset markets, such as those for Counter-Strike 2 weapon skins, are characterized by their small size, high volatility, and strong influence from community discussions and platform rules. These unique properties make them challenging for traditional quantitative trading models but offer an excellent environment for studying how large language models (LLMs) can convert unstructured textual information into actionable trading decisions. CSTrader is a new multi-agent framework developed for language-grounded trading in this specific market. The system first aggregates various signals from multiple sources, then employs specialized agents for tasks like technical analysis, assessing liquidity, monitoring events, and interpreting sentiment (including reversed sentiment). Finally, risk control, transaction friction, and portfolio management agents are applied to generate buy, sell, or hold decisions, accounting for realistic trading costs. Evaluated in a live-like environment using real CS2 data from a highly volatile period, CSTrader consistently outperformed both a declining market index and simpler single-prompt LLM baselines, achieving up to a 7.58% cumulative return with controlled risk. Ablation studies highlighted the critical roles of liquidity, reversed sentiment, and transaction friction agents in transforming noisy language signals into stable profits, suggesting these niche, language-driven markets are valuable benchmarks for future language-to-action research.