The technical round focused intensely on designing an ML model to predict short-horizon equity returns, diving deep into feature selection challenges with multicollinearity and the shortcomings of traditional k-fold cross-validation for financial time series. There was also an emphasis on evaluating model performance with real transaction costs. Surprisingly, the complexity of the questions mirrored some of the company-specific prompts I found on PracHub during my prep, which helped me feel a bit more prepared. Overall, the process was rigorous, and the lack of an offer was disappointing after such challenging discussions.

Design an ML model to predict short-horizon equity return signals from a high-dimensional, noisy feature set. Walk through how you'd handle feature selection under heavy multicollinearity, why standard k-fold cross-validation breaks down on non-stationary financial time series and what walk-forward / purged-CV setup you'd use instead, how you'd evaluate the model out-of-sample with realistic transaction costs, and how you'd detect when alpha decay justifies a full retrain vs. an online update.