longbridge-ml-strategy

Walk-forward machine-learning framework for stock direction prediction. Fetches historical OHLCV data, engineers technical features, trains a rolling classifier (Random Forest or Gradient Boosting), generates probabilistic buy/sell signals, and evaluates backtest performance.

Response language: match the user's input language — Simplified Chinese / Traditional Chinese / English.

Dependencies

Requires: scikit-learn, pandas, numpy (usually pre-installed). Optional: xgboost or lightgbm for gradient-boosting models. If unavailable, fall back to a simpler logistic-regression model.

When to use

• User asks for ML-based prediction, rolling model training, feature-importance analysis, or AI-driven entry/exit signals for a single stock.
• Triggers: "用机器学习预测 TSLA 涨跌", "NVDA random forest strategy", "walk-forward backtest AAPL".

Workflow

1. Fetch 504 daily candles (≈ 2 years): longbridge kline <SYMBOL> --period day --count 504 --format json

2. Feature engineering (compute on rolling windows):

   • MACD line and signal (EMA12 − EMA26, signal EMA9)
   • RSI-14
   • Bollinger Band width: (upper − lower) / mid, window 20
   • Volume change rate: (vol_t − vol_{t-5}) / vol_{t-5}
   • 5-day price momentum: (close_t / close_{t-5}) − 1
   • Label: 1 if close_{t+5} > close_t × 1.01, 0 if < close_t × 0.99, else drop

3. Walk-forward training:

   • Training window: 252 days; retrain every 60 days
   • Model: RandomForestClassifier(n_estimators=100) or GradientBoostingClassifier
   • Predict probability for the current bar

4. Signal generation:

   • prob > 0.60 → 模型上涨概率偏高 / Model upside probability elevated
   • prob < 0.40 → 模型下跌概率偏高 / Model downside probability elevated
   • Otherwise → 模型无方向性预测 / No directional signal from model

5. Backtest metrics (on out-of-sample predictions):

   • Win rate (% correct directional calls)
   • Profit factor (gross profit / gross loss)
   • Annualised Sharpe ratio (assuming daily rebalance)
   • Max drawdown

6. Feature importance: rank top-5 features by mean decrease in impurity.

Run longbridge kline --help to confirm flag names before calling.

CLI

longbridge kline --help

longbridge kline <SYMBOL> --period day --count 504 --format json

Output

Output: current signal box → backtest summary table → feature importance list → caveats (past performance, data snooping). Cite Longbridge Securities / 数据来源：长桥证券 / 數據來源：長橋證券.

以上内容仅供参考，不构成投资建议。投资决策请结合自身风险承受能力独立判断。 The above is for reference only and does not constitute investment advice. Investment decisions should be made based on your own risk tolerance.

Error handling

MCP fallback

When the CLI is unavailable, fall back to the MCP server. Discover available tools from the MCP server's tool list at runtime.

Related skills

• longbridge-kline — OHLCV data source
• longbridge-volatility-strategy — vol regime as an additional feature
• longbridge-multifactor — cross-sectional factor signals complement

File layout

longbridge-ml-strategy/
└── SKILL.md