Othis
Jun 2026 – Aug 2026
Quantitative Analyst Intern · Vienna, Austria
- Built the covariance forecasting engines for a cross-asset factor allocation strategy. Fifteen estimators span two families: GARCH/DCC models (with linear and nonlinear shrinkage) and HAR-class models built on realized measures from one-minute bars.
- Wrote the evaluation pipeline that scored every engine on two tracks: (1) QLIKE and Frobenius loss with Diebold-Mariano tests; (2) weekly and monthly portfolio rebuilds measured on realized volatility, Sharpe, and drawdown with stationary bootstrap significance.
- Ran the study across three panels: 138 daily factor returns, 100 US single stocks at one-minute frequency, and 3,500 daily equity names.
- Replicated Engle-Ledoit-Wolf (2019) on the largest panel, extending their dimension ladder from a Monte Carlo to real data.
- Designed two in-house extensions: a gradient-boosted correction to the HAR variance forecast, and a reframing of the DCC dynamic layer as a continuous intensity rather than a binary model choice.
- Primary finding: QLIKE penalizes a forecast's scale, while a minimum-variance allocator sees only shape (the two metrics rank engines differently). At a 21-day horizon, a martingale realized covariance was worse than its benchmark on QLIKE in 92% of windows yet built the lowest-volatility portfolio of any engine; scale accounted for 60% of that QLIKE gap in the median window (75% at five days).
Technologies: Python, NumPy, pandas, arch, nonlinshrink, skfolio, cvxpy, Numba.