Assessing Machine-Learning Strategies and Their Live Trading Risks

Summary

This Q&A compilation collects practitioner questions about developing, combining, and deploying quantitative strategies. Topics include variation between XGBoost rolling-training backtests, selecting and evaluating factors and models, overlapping holdings across strategies, assessing style contributions to portfolio risk, and deciding whether a strategy has weakened by comparing live excess returns and drawdowns with expectations.

The document also raises practical design questions: whether longer backtests establish robustness, how to use large libraries of ordinary and high-frequency factors, and how to move from factor discovery through model training and simulation to live trading. Its only direct response is an announcement that a more continuous course covering the full workflow was planned. It does not answer most questions or provide a specific evaluation method, experimental evidence, or deployment thresholds. As a result, it is useful chiefly as an inventory of issues researchers should address, rather than as a guide that resolves them.

Key ideas

• Rolling XGBoost training can produce materially different backtests under seemingly identical settings, prompting questions about randomness and evaluation.
• Strategy combination raises issues such as overlapping stock selections and styles whose risk contribution may exceed their return contribution.
• The questions emphasize monitoring live excess returns and drawdowns against backtest expectations when judging strategy health.
• Robustness, factor selection, model choice, and the transition from simulation to live trading are recurring concerns.
• The document provides no detailed answers to most questions and reports no empirical results.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.