Quantitative Research · Financial Data · C++/Rust · High-Performance Computing
Data Science undergraduate at HKUST(GZ) and Quantitative Investment Modeling Research Intern. My work spans A-share Level-2 market microstructure, leakage-resistant financial time-series evaluation, scalable data pipelines, C++/Rust systems, and HPC.
I prefer results that remain explainable under realistic information boundaries: explicit sample definitions, point-in-time alignment, out-of-time validation, independent correctness checks, and benchmark-driven performance work.
current technical questions
My current work is concentrated in three connected areas: extracting interpretable structure from high-frequency market data, evaluating financial models under causal information constraints, and building systems that keep large-scale research correct and efficient.
How do execution topology, active/passive order structure, direction continuity, duration and price paths differ even when the same fixed volume is traded?
How should signals and forecasting models be evaluated when information arrives sequentially, regimes change, features overlap, and apparently strong effects may disappear out of time?
How can data loading, deterministic replay, numerical workloads and cluster execution be made faster without weakening correctness, reproducibility or observability?
industry and technical training
Current research uses A-share Level-2 and minute data to study fixed-volume execution structure, event response, information availability, stability and redundancy. Public descriptions intentionally omit proprietary signals and employer code.
I also maintain multi-year market-wide research pipelines and improved transposed cross-day DataLoader access for roughly 4× faster data loading.
Worked on C++/CUDA inference optimization on NVIDIA A800 clusters with TensorRT, low-precision inference, profiling, MPI and Slurm. In one internal UniFOLM-WMA task, runtime fell from about 350s to 48s and GPU memory from about 12.4GB to 4.8GB.
Team results: First Prize in 2026 and Second Prize in 2025.
Coursework included mathematical science, quantum physics, quantum electronics / information and circuit theory.
selected academic work
Market-wide research on how fixed-volume bars are completed: execution topology, active/passive order concentration, direction switches, path efficiency, duration, sparse gaps and post-confirmation quote response.
Evaluation principle: mechanism evidence is kept separate from tradable-alpha claims until point-in-time, holdout and incremental tests are completed.
Co-author. Built terabyte-scale financial time-series preprocessing and research infrastructure. Internal benchmarks achieved up to 13× faster data cleaning and 6× faster statistical aggregation.
Research on evidence aggregation, claim-level verification, calibration, point-in-time provenance and temporal grounding under noisy, changing financial information.
Current stage: framework/evaluation design, model/API and compute preparation, and automated experiment-environment setup; no completed-result claim.
Independent diffusion-model research on training-free controllable generation, covering problem formulation, method design, theoretical analysis, implementation, controlled/ablation experiments, manuscript writing and a full peer-review cycle.
interviewable systems work
Implemented deterministic event replay ordered by information availability and an L2 price-level order-book consumer. An independent Python oracle checks cross-language consistency; the project records 144 runs and 11,638 lines of book-state output, with sanitizer/conformance CI and controlled benchmarks.
Leakage-safe rolling out-of-sample comparison of Historical Simulation, Student-t GARCH/GJR-GARCH, realized-measure quantile regression, MLP and GRU for 1% / 5% / 10% conditional VaR. Development choices are frozen before the final test; GJR-GARCH was more stable across the three tail quantiles while small neural models showed no consistent advantage.
depth before labels
C++: long-term user; strong command of C++17/20 and earlier modern C++, partial C++23 knowledge, and ongoing tracking of C++26. RAII, move semantics, templates/concepts, STL/ranges, lambdas, coroutines, concurrency, CMake, sanitizers, benchmarking and profiling.
Rust: practical experience with ownership/borrowing, enum/match, traits/generics, Result/Option and Cargo; used in reliable state and event-replay components.
Python / ML: NumPy, Polars/pandas, SciPy, scikit-learn, PyTorch, LightGBM, XGBoost, SQL.
Quant: Level-2/minute data, market microstructure, factor/neural-feature research, TS/CS IC, walk-forward/OOT, stability/decay/redundancy and leakage audits.
Systems / HPC: Linux, Git, Docker, CUDA, TensorRT, MPI, Slurm, parallel I/O, memory mapping, GPU/CPU profiling, CPU affinity and NUMA/cache locality.
peer-reviewed work
foundation & distinctions
BSc, Data Science and Big Data Technology · 2023-2027 (expected)
CGA 3.807/4.3 · Major CGA 3.837 · Dean's List ×3.
Selected foundations: algorithms and data structures, linear algebra, calculus, statistics, machine learning, databases, cloud/big-data systems and optimization.