Aegis Equity: A Falsification-Driven Framework for Quantitative Trading and Model Risk Management in Investment Banking

The proliferation of machine learning tools has made it straightforward to construct a quantitative trading strategy that generates impressive historical performance. The same tools have made it equally straightforward to construct a strategy that merely appears to perform well, one whose apparent edge is an artifact of overfitting, data leakage, or multiple testing rather than genuine predictive skill. Distinguishing between the two is the central methodological challenge of applied quantitative finance, and the primary contribution of this thesis. This research constructs, validates, and deliberately attempts to falsify Aegis Equity, a dual LightGBM classifier trained on 14 endogenous price-and-volume features, applied to the NSE 500 universe with a Beta-Switch regime filter that dynamically allocates capital between stock-picking, index exposure, and defensive cash. The system is evaluated through a pre-registered three-gate falsification protocol in which all thresholds, hypotheses, and decision rules are locked before the out-of-sample window (January 2019 to October 2023) is examined. Gate 1 confirms the machine learning models possess statistically significant trade-level predictive skill. Gate 2 formally rejects naive portfolio translation, demonstrating that trade-level skill alone is insufficient to generate portfolio-level outperformance during systemic market shocks. Gate 3 confirms that the Beta-Switch regime filter resolves this failure: the validated architecture achieves a Sharpe ratio of 2.2461 against a benchmark Sharpe of 0.7784, clearing a Bonferroni-corrected block bootstrap threshold of p < 0.0166 across all three tested block sizes. The thesis documents five backtest construction failure modes, narrative inflation, feature-level data leakage, execution logic bugs, temporal scope violations, and statistical misalignment, each detected through artifact-level verification rather than statistical flagging, along with a sixth discarded test whose invalidity is documented rather than hidden. These findings are mapped onto the model governance practices of investment banking institutions, demonstrating that the three-gate protocol is a practical, working implementation of the independent validation principles mandated by regulatory frameworks including SR 11-7.