Skip to content

CSTrader: A Testbed for Language-Grounded Trading in a Community-Driven Virtual Asset Market

🕒 Published (v1): 2026-06-30 10:40 UTC · Source: Arxiv · link

Why this paper was selected

Novel agent testbed with grounded language-trading environment; valuable for harness eval

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CSTrader is a three-tier multi-agent LLM framework that treats the Counter-Strike 2 skin market as a testbed for language-grounded trading, integrating price data, Reddit sentiment, and official game-update events to produce buy/sell/hold decisions under realistic transaction frictions. Evaluated on a highly volatile real-market period (Sep 25–Nov 15), it achieves up to 7.58% cumulative return against a market index that fell 15.62%. The paper's primary contribution is the evaluation environment itself, not just the agent policy.

Problem

Traditional quantitative models and single-prompt LLM approaches fail on niche, community-driven asset markets (e.g., CS2 weapon skins) because such markets are dominated by unstructured social text, platform-specific rules (trade-lock periods, ~2% fees), thin liquidity, and herd dynamics—none of which standard stock-trading frameworks model. Existing backtesting environments also lack realism (lookahead bias, no friction modeling).

Method

CSTrader is organized in three tiers:

  1. Information Perception Layer: Streams daily OHLCV data (Steam JSON API), Reddit community posts (sliding time-window alignment), and official Steam/CS2 patch news into standardized inputs.

  2. Reasoning Layer (5 specialized agents): Each outputs a discrete Bullish/Bearish/Neutral signal with a natural-language justification:

  3. Technical Agent: EMA, Bollinger-band mean-reversion, RSI-type momentum, volatility regimes, support/resistance.
  4. Sentiment Agent: Reddit sentiment per skin (standard polarity).
  5. Reversed Sentiment Agent: Contrarian—inverts community hype to exploit bubble/overreaction dynamics.
  6. Liquidity Agent: Combines trading volume with Reddit engagement to flag "liquidity traps" (high listed price, low real demand).

  7. Event Agent: Classifies Steam patch notes as bullish/bearish/neutral per skin.

  8. Operation Layer:

  9. Risk Control Agent: Aggregates Reasoning Layer signals + portfolio state → proposed position ratio, upweighting defensive signals in uncertain periods.
  10. Transaction Friction Agent: Hard-codes the 2% sell fee; blocks trades where expected profit is negative after costs.
  11. Portfolio Management Agent: Final decision-maker combining position ratio, friction check, cash balance, and a short decision memory → discrete BUY/SELL/HOLD + trade size.

Six frontier LLM backbones are evaluated as drop-in replacements (DeepSeek-V3.2, Qwen-Max, Moonshot-v1, Gemini-3-flash-preview, GPT-5-mini, Claude-sonnet-4, Grok-4.1-fast-reasoning) within the same fixed framework interface.

Key Contributions

  • A realistic evaluation environment for language-grounded trading that streams heterogeneous textual and numerical signals while enforcing real-world frictions (fees, trade-lock), instantiated on CS2 skins but designed to generalize to NFTs and collectibles.
  • A structured multi-agent LLM baseline (CSTrader) decomposing perception, multi-dimensional reasoning, and risk-aware portfolio control into interacting prompt-engineered agents with explicit natural-language rationales.
  • Empirical demonstration that liquidity filtering, reversed sentiment (contrarian), and explicit transaction friction modeling are the three most critical ablation components for stable positive returns in a niche market.
  • A backbone comparison showing that both reasoning quality and risk preference of the underlying LLM measurably affect performance within a fixed framework.

Results

  • vs. market index (fell 15.62%): All strong-backbone CSTrader configs achieve positive CR:
  • Claude-sonnet-4: CR = +7.58%, SR = 0.49, MDD = 25.9%, AV = 79.81%, \(\alpha\) = 66.86%, \(\beta\) = 0.47
  • Qwen-Max: CR = +3.83%, SR = 0.42, MDD = 17.04%, AV = 38.86%, \(\alpha\) = 28.47%, \(\beta\) = 0.20 (best risk-adjusted)
  • Moonshot-v1: CR = +4.50%, \(\alpha\) = 43.62%
  • GPT-5-mini: CR = +4.29%, \(\alpha\) = 54.25%
  • DeepSeek-V3.2: CR = −2.89%, Grok-4.1: CR = −6.48%
  • vs. single-prompt LLM baseline: Simple prompts yield CR = −11.50%, SR = −0.75%; full CSTrader (Qwen, best combination) yields CR = +3.83%
  • Ablation (Qwen-Max backbone):
  • Technical + Friction only: CR = +1.62%
    • Reversed Sentiment + Liquidity + Friction: CR = +5.94%
    • Sentiment (non-reversed): CR = −10.77% (worse than market in isolation)
  • No Friction (same agents): CR = +21.96%, but MDD = 31.06%, AV = 95.04% (unrealistic)
  • Event Agent adds limited or negative incremental value; Official Steam news is priced in quickly.

Limitations

  • Single niche market (CS2 skins, one liquidity-filtered item subset); transferability to regulated or larger markets is unverified.
  • Evaluation window is short (~7 weeks) and concentrated around one major shock event ("Trade Up" update, Oct 25).
  • No order-book dynamics, slippage, or execution latency modeling beyond a flat 2% fee.
  • Agents are prompt-engineered only; no fine-tuning or systematic hyperparameter/architecture search.
  • No classical quantitative (e.g., momentum factors, mean-reversion strategies) or reinforcement-learning baselines for comparison.

Relevance to Harnesses / Meta-Harnesses

CSTrader is directly relevant as a domain-specific multi-agent harness: it instantiates a fixed three-tier orchestration framework in which specialized agents (technical, sentiment, liquidity, event, risk, friction, portfolio) are composed into a pipeline with a standardized interface, allowing any LLM backbone to be swapped in without changing the harness structure. The evaluation environment itself is designed as a meta-harness—an agent-agnostic wrapper that enforces real-world constraints (fees, trade locks, streaming data) and enables fair comparison across agent architectures. This pattern—separating the harness from the policy, then using the harness to benchmark multiple instantiations—mirrors the design philosophy of agent evaluation frameworks in the broader harness literature. The ablation methodology (systematically adding/removing sub-agents) is also a practical template for decomposing multi-agent harness contribution.