Quantitative Multi-Factor Timing

01 / Context

The problem

Technical timing strategies often depend on fixed indicator parameters and a single model, making them fragile when market regimes change. The project asked whether feature engineering, ensemble learning and clustering-based parameter updates could produce a more robust signal process.

02 / Approach

What I did

• Engineered 14 groups of technical-indicator features and encoded Bollinger Band, ADX and EMA signal rules for systematic backtesting.
• Built a stacking ensemble combining XGBoost, LightGBM, Random Forest and LSTM, then optimised decision thresholds.
• Ran a two-layer grid search across Bollinger periods and standard-deviation multiples.
• Used K-means clustering of SuperTrend factor performance to support dynamic parameter updates.
• Developed and backtested a KNN classification strategy using Manhattan distance across multi-dimensional indicators.

03 / Outcome

What came out of it

• Backtested the classification strategy on 20,000 five-minute steel-rebar bars.
• Produced a documented research report covering signals, model design and backtest outcomes.
• Combined time-series forecasting, tree-based learning and regime-aware parameter selection in one research workflow.

04 / Artifacts

Reports & code