Mehmet Demir Güven

Zürich, Switzerland | dgueven@student.ethz.ch | GitHub | LinkedIn

Education

ETH ZürichZürich, Switzerland

Bachelor of Science in Computer ScienceSep 2026 - Expected Aug 2029

Coursework: Linear Algebra, Data Structures and Algorithms, Discrete Mathematics, Introduction to Programming

Independent ResearchZürich, Switzerland

Parameter Symmetry in Sine-Network Implicit Neural Representations (SIREN-GAP)Jul 2026 - Present

Proved that neuron permutations, sign flips and π bias shifts generate the complete function-preserving symmetry group of generic sine networks with 1 and 2 hidden layers
Showed with a Shapley-value factorial experiment that this group alone accounts for 79.1 of the 80.4 percentage points of MNIST accuracy lost when INRs start from independent rather than shared initializations
Designed an exactly invariant weight-space classifier via a (cos b, sin b) bias encoding that closes 91.7% of the gap at equal parameter count, against 62.8% for the best alignment method
Released the method as phasorkit, a pip-installable PyTorch library with invariant layers and an invariance certifier; manuscript in arXiv moderation

Low-Rank Confounding and Partial Identification of Cross-Asset Price Impact (XID)Jul 2026 - Aug 2026

Proved that the confounding gap between estimated and true cross-impact matrices has rank at most K + rank(B), verified in a preregistered simulation with 30 assets, 3 factors and 10M observations
Derived the execution-cost error in closed form: an equal-weight index basket is mispriced by 54% of its true cost, a dollar-neutral basket by 0%

Reconstructing the Unlisted FLEX Options Market from Clearing Data (FLEX)Aug 2026 - Sep 2026

Built a validated pipeline turning OCC daily clearing reports into a 13-month panel of the unlisted FLEX market: 269 activity dates, 1,691 underlyings and 9.06M series-days, reaching 73.0M contracts and $596B of mark value
Showed strike-based FLEX detection has precision 0.67 and recall 0.34 and that leg-by-leg counting overstates a butterfly 49x; a placebo-tested fixed-effects regression finds no next-day predictability

Forecasting Run-to-Run Variance in Reinforcement Learning Post-Training of LLMs (caliper)Aug 2026

Derived a covariance recursion and adjoint backward pass that predicts the across-seed standard deviation of a model's pass rate from one stored training run, reaching 1.11x median error over 192 prospective settings
Showed the reward signal drives variance contraction: replacing rewards with coin flips turned a divergence that shrank to 0.14x into one that grew 90.5x; reported where the forecast fails on Qwen2.5-0.5B and 7B

TÜBİTAK National Informatics Olympiad - Second-round qualifier, 33rd National Science OlympiadsDec 2025

Qualified as 1 of 511 from 18,463 students nationwide (top 3%) for the TÜBİTAK summer school and final

TÜBİTAK National Mathematics Olympiad - Second-round qualifier, 32nd National Science OlympiadsDec 2024

Advanced to the proof-based national round: number theory, combinatorics, algebra and geometry

Languages: Python, C++, Java, Rust, TypeScript, Swift, SQL, Bash, LaTeX

ML and Scientific: PyTorch, MLX, NumPy, SciPy, pandas, Polars, scikit-learn, Hugging Face Transformers

Quant Methods: Purged walk-forward CV, Clark-West tests, FDR control, deflated Sharpe ratio, stationary bootstrap

Algorithms: Dynamic programming, graph algorithms, advanced data structures

Mathematics: Probability, matrix analysis, group theory, convex optimization

Spoken: Turkish (native), German (C2, DSD II), English (C1)