Built a sequence-based deep learning subsystem for real-time trade-size estimation, designing asymmetric loss functions over right-censored heavy-tailed distributions and capturing 17% incremental unmapped institutional liquidity.
Designed gradient-boosted quantile regression models with pinball loss for calibrated prediction intervals, adding isotonic post-processing to eliminate quantile crossing and reduce pricing-engine sensitivity to anomalies by 22%.
Modernized offline backtesting infrastructure with walk-forward validation and point-in-time feature construction to prevent lookahead bias; containerized (Docker) pipelines improving validation throughput by 9%.
Built a sequence-based deep learning subsystem for real-time trade-size estimation, designing asymmetric loss functions over right-censored heavy-tailed distributions and capturing 17% incremental unmapped institutional liquidity.
Designed gradient-boosted quantile regression models with pinball loss for calibrated prediction intervals, adding isotonic post-processing to eliminate quantile crossing and reduce pricing-engine sensitivity to anomalies by 22%.
Modernized offline backtesting infrastructure with walk-forward validation and point-in-time feature construction to prevent lookahead bias; containerized (Docker) pipelines improving validation throughput by 9%.