[GitHub] wonglaitung/fortune [GitHub] wonglaitung/fortune：金融资产量化分析与交易系统

TL;DR

• Human-machine hybrid intelligence system targets Hong Kong stocks with 81.22% accuracy.
• "False breakout long" strategy achieves 87% win rate using multi-period predictions.
• CatBoost model utilizes 1,023 features for individual stock analysis, scoring 90 in validation.
• Automated via GitHub Actions, delivering daily trading signals through email notifications.

Key Data

Deep Analysis

The financial technology sector is saturated with "AI trading" projects that promise the moon but deliver nothing more than overfitted backtests. This GitHub project, however, cuts through the noise with a refreshingly pragmatic approach: it doesn't try to replace the human, it tries to augment them. The core philosophy here is "Human-Machine Hybrid Intelligence," a buzzword-laden concept that actually translates to a sensible architecture. By fusing the reasoning capabilities of Large Language Models (LLMs) with the numerical precision of machine learning, the system acknowledges a fundamental truth—markets are not just numbers; they are narratives.

Let's address the elephant in the room immediately: the 81.22% accuracy rate for 20-day predictions. In the world of quantitative finance, numbers this high usually scream "data leakage" or "overfitting." However, the project mitigates this skepticism by employing Walk-forward validation. This is not your standard train-test split where a model memorizes history; Walk-forward validation simulates real-time trading by rolling the training window forward. If the 90-point validation score holds up out-of-sample, this isn't just a script; it's a legitimate alpha generator. The 87% win rate on the "false breakout long" strategy is particularly intriguing. False breakouts are the bane of retail traders, trapping emotional buyers at the top. A model that systematically identifies and fades these traps is essentially monetizing the inefficiencies of human psychology.

Technically, the stack is robust. The use of CatBoost is a smart choice for tabular financial data, often outperforming deep learning models on structured datasets. But the real innovation lies in the "Network Cross Features." Most retail quants stop at RSI and MACD. This project incorporates network theory, analyzing how stocks respond to market signals as a correlated group. This addresses the "convergence" problem where individual stocks simply move in lockstep with the index, rendering standard analysis useless. By differentiating how specific stocks react to broader market network signals, the model captures a layer of nuance that standard technical analysis misses.

The integration of HMM (Hidden Markov Models) for market state recognition is another layer of sophistication. Markets have distinct personalities—trending, mean-reverting, or chaotic. A strategy that works in a bull market dies in a range-bound one. HMM allows the system to identify these regimes and adjust thresholds dynamically. This is where the "dynamic threshold mechanism" shines. Instead of a static "buy when score > 0.8," the system adapts to the volatility environment, a crucial feature for survival in the volatile Hong Kong market.

However, the reliance on GitHub Actions for automation is a double-edged sword. It democratizes access, allowing users to run institutional-grade strategies for free. But it introduces latency and dependency on external infrastructure. For high-frequency needs, this is useless; for the daily rebalancing this system targets, it is perfectly adequate.

Ultimately, the project’s value proposition is the removal of "rigidity." Traditional quant strategies are brittle; they break when market regimes shift. Pure human analysis is slow and biased. This system sits in the sweet spot, using ML to process the 1,023 features no human could track, and LLMs to contextualize the output. It is a tool built for the modern trader who understands that the future isn't man vs. machine, but man directing the machine.

Industry Insights

1. Feature Engineering is the New Alpha: With 1,023 features including network cross-data, the industry is moving beyond simple technical indicators to complex, relational data structures.
2. Regime Awareness is Mandatory: The use of HMM for market state detection highlights a shift from static strategies to adaptive, context-aware trading systems.
3. Validation over Optimization: The emphasis on strict Walk-forward validation signals the end of "curve-fitting" culture; robustness is now prioritized over hypothetical returns.

FAQ

Q: How does the "Human-Machine Hybrid" approach actually work in practice?
A: The system automates data processing and pattern recognition via ML, while LLMs likely interpret context, leaving the final execution and risk management to the human investor.

Q: Is an 81% prediction accuracy realistic for long-term trading?
A: While high, the metric is likely directional accuracy for a 20-day trend; profitability depends on risk-reward ratios and slippage, not just hit rate.

Q: What makes the "Network Cross Features" different from standard indicators?
A: Unlike standard indicators that look at stocks in isolation, network features analyze how a stock's price action correlates and interacts with the broader market network structure.

TL;DR

• 港股量化系统融合大模型与机器学习，恒指20天预测准确率达81.22%。
• 个股分析整合1023个特征，Walk-forward验证评分90，有效防止过拟合。
• 创新引入“网络交叉特征”，结合社区情绪与市场信号，提升策略胜率。
• 全流程自动化运行，特定策略如“假突破做多”胜率高达87%。

核心数据

深度解读

在量化交易这个早已拥挤不堪的赛道上，大多数开源项目要么是徒有其表的“玩具”，要么是过度拟合的“回测战神”。但这个项目却极其罕见地展示了一种老练的实战主义——它不迷信单一模型的万能，而是通过“人机混合智能”的架构，承认了市场的混沌与模型的局限。

首先，那个高达81.22%的20天预测准确率，初看令人咋舌，细想却不仅是算法的胜利，更是“特征工程”的降维打击。项目最值得玩味的技术创新在于构建了“网络交叉特征”。传统的量化模型往往死盯着K线和财务报表，却忽略了社交媒体时代市场情绪的传染性。将网络社区的特征与市场级特征交叉，实际上是在捕捉“羊群效应”的数字化足迹。这种做法极其敏锐，它不再把市场看作冷冰冰的数字集合，而是一个充满噪音和情绪的有机体。这比单纯堆砌LSTM或Transformer层要高明得多。

其次，项目采用Walk-forward验证并给出90分评分，这才是专业选手的底色。市面上太多模型死在“过拟合”这道坎上——在历史数据上拳打巴菲特，一上实盘就亏损累累。该项目坚持使用Walk-forward这种极其严苛的验证方式，说明开发者深知金融数据的非平稳性，拒绝自欺欺人。配合动态阈值机制应对极端行情，这套系统展现出了极强的风控意识，而非单纯的收益赌博。

然而，必须泼一盆冷水：87%的胜率数据具有极强的迷惑性。胜率不等于盈亏比，高频交易中常见的“赚小钱亏大钱”陷阱依然存在。虽然系统引入了大模型做推理，但在极端的流动性危机面前，无论是大模型还是HMM状态识别，都可能瞬间失效。这个项目的真正价值，不在于那个诱人的准确率数字，而在于它提供了一套可复用的、将非结构化信息（网络情绪）结构化的高效框架。它试图证明，在AI时代，量化交易的核心竞争力已从单纯的算力军备竞赛，转向了对异构数据的深度理解与融合能力。

行业启示

1. 量化竞争已从单纯的技术指标计算转向“另类数据”挖掘，网络情绪与市场信号的交叉融合将成为Alpha新来源。
2. 模型验证的严谨性决定生死，Walk-forward验证应成为量化项目的标配，杜绝静态回测带来的虚假繁荣。
3. “人机混合”并非倒退，大模型负责逻辑推理、机器学习负责精度预测的分工模式，将重塑金融分析流程。

FAQ

Q: 这个系统适合完全没有编程基础的散户直接使用吗？
A: 不适合。虽然实现了自动化，但需配置环境、API及理解策略逻辑，具备一定技术背景者方能驾驭。

Q: 81.22%的预测准确率是否意味着稳赚不赔？
A: 绝非如此。准确率仅代表方向判断，不包含盈亏比；且实盘滑点、流动性变化均可能导致实际收益大幅缩水。

Q: 什么是“网络交叉特征”，为何它是本项目的亮点？
A: 指将网络社区的情绪特征与市场交易数据结合。它能捕捉散户情绪对股价的非线性影响，解决了传统量化忽视“人心”的缺陷。