Quantitative Research Internship

From indicator rules to model features

Each strategy was decomposed into observable inputs rather than left as a hard-coded buy/sell rule. Bollinger periods and standard-deviation multiples were varied, while ADX and EMA contributed trend-strength and direction context. The resulting matrix supported both simple classification and ensemble learning.

• Indicator parameters were separated from model parameters so each could be evaluated independently.
• Signal rules were preserved alongside raw indicator values to retain interpretability.
• Five-minute bars created a high-noise target, making threshold selection part of the model rather than an afterthought.

Stacking and adaptive parameters

A heterogeneous stack combined tree models, an LSTM and KNN. The first-stage models captured different nonlinearities, while a meta-learner combined their outputs. A two-layer grid search selected Bollinger parameters, and K-means clustering of SuperTrend performance supported later parameter updates when the factor regime changed.

• Tree models handled tabular nonlinearities and interactions efficiently.
• LSTM introduced a temporal representation but required stricter sequence construction.
• Clustering was used as a regime-aware parameter policy, not as a direct price predictor.

Backtest discipline

The workflow produced a research report as well as code. The analysis intentionally separated in-sample fitting from the historical evaluation output and documented where the model was fragile. The current limitation is that production-grade transaction costs, slippage and live execution constraints would need a more explicit simulation layer.