Quantitative Finance Researcher

Representation learning for markets: a 4.9M-parameter time-series transformer that learns temporal and cross-asset structure across 499 stocks.

• Trained a 4.9M-parameter financial time-series transformer on 24.8M observations across 499 tickers, 12 features, and 4,149 daily bars.
• Designed a temporal and cross-asset representation-learning architecture extending temporal neighborhood coding with graph convolutions and forward-aligned objectives using 120-day historical correlations or learned attention; submitted to ICLR 2027.
• Benchmarked 9 representations across 21 targets, 3 horizons, and 3 seeds on 37,500 embargoed windows; at 21 days, improved relative-volatility and beta rank IC by 0.110 and 0.090 over the strongest baselines, respectively, with under 0.015 seed variation.