Strategy Hallucination
The model emits plausible trading logic that looks quantitative but has no market rationale, no statistical grounding, or invalid assumptions.
It creates code that passes syntax checks while smuggling in false causality.结构化洞察
交易软件的代际演进：从 MT4 到 LLM
四代交易软件的结构化对比：指标时代平台、策略框架、系统管线和 LLM 提示词生成代码工作流。
四代交易软件
关键变化不仅是语言或用户体验。每一代都改变了交易者用来表达、测试和实施策略思路的核心抽象。
| 维度 | Gen 1: Indicator Era~2005-2012 | Gen 2: Strategy Era~2012-2018 | Gen 3: System Era~2018-2023 | Gen 4: LLM Era2023-present |
|---|---|---|---|---|
| Time period | ~2005-2012 | ~2012-2018 | ~2018-2023 | 2023-present |
| Representative products | MT4 / MQL4 | Backtrader, freqtrade, vnpy | QuantConnect, WorldQuant BRAIN | ChatGPT + brokerage API |
| Core abstraction | Modular indicators | Packaged strategy logic: entry, exit, sizing | Feature engineering pipeline | Natural language to code |
| Typical workflow | Drag indicators, set conditions, backtest | Write strategy code, optimize parameters, backtest | Build feature pipeline, train, validate, execute | Prompt, generate strategy, backtest |
| Main progress | Indicators moved from books and forums into reusable components | Complete strategies became packageable, shareable, and reproducible | The pipeline became the product, with standardized validation | Generation is fast and the entry barrier is low |
| Typical trap | Indicator worship and holy-grail thinking | Parameter overfitting: genetic search finds coincidence, not robustness | Data leakage and survivorship bias | Strategy hallucination and backtest overfitting blindness |
| Failure root cause | Rules are manual and lack systematic validation | The optimization target is backtest return, not robustness | Model complexity hides data-quality problems | The model skips infrastructure accumulated across the first three generations |
LLM 时代的两种失败模式
LLM 降低了生成策略代码的成本，但并未消除对验证基础设施的需求。
Backtest Overfitting Blindness
The model treats a profitable backtest as validation and misses leakage, parameter mining, unstable regimes, or survivorship bias.
It accelerates curve fitting because generation speed multiplies untested variants.LLM 经常生成的 Z-score 模式
spread = asset_a.close - hedge_ratio * asset_b.close
z_score = (spread - spread.mean()) / spread.std()
if z_score > 2:
short(asset_a)
long(asset_b)
elif z_score < -2:
long(asset_a)
short(asset_b)模型 + 工具
模型不是交易系统
LLM 在被工具包裹时最有用。一个严肃的交易技术栈包括数据验证、定价引擎、执行逻辑、风控、监控和部署纪律。
Jane Street 风格的系统展示了这一模式：模型只是更大工具链中的一层，而非整个产品。
StratCraft 定位
面向第2代和第3代开发者的基础设施
StratCraft 不自称为第5代。它为策略框架和系统管线开发者提供本地 C++ 级别的回测性能、插件隔离和可重复的验证工作流。