01 Home02 Research03 Projects04 Notes05 About06 Lab
Home/Projects

Systems, evaluations
and useful failures.

Three current projects, each with one figure, a plain account of what I did and where the evidence stops. Earlier research follows below.

01 · Independent system · 2026 · ongoing

Helix

A quantitative research and paper-trading system built around one shared signal core. It takes a signal from data provenance through backtests, portfolio risk checks and validation gates to a simulated paper-trading loop, across US, China A-share and Hong Kong equities.

My role
Design, implementation and validation
Builds on
Qlib for one ML sleeve; a factor engine vendored from HKUDS Vibe-Trading (MIT)
Status
Research and paper workflows; real-capital execution gated

Limits: the paper track record includes replayed days, and survivorship inflates absolute backtest returns by about five points a year in my own check, so no absolute returns are quoted.

Read the case: question, role, evidence, limits →
Helix · Research terminal · snapshot 27 Jul 2026
The Helix research terminal: a four-strategy book with its combined equity curve, signal-parity and risk checks passing, and the deflated-Sharpe validation failing.
Helix · Exhibition
The Helix exhibition page: 23 live browser experiments in five wings.
Validation gate · this snapshot
Signal parity · passRisk · passDeflated Sharpe 0.21 · fail
The system rejects its own book.

Real screenshots of the Helix web front. The terminal shows a saved snapshot (27 Jul 2026, 50 liquid US stocks): an in-sample backtest whose combined book fails Helix’s own deflated-Sharpe gate. Its numbers are not performance claims.

02 · Independent evaluation · Jul – Sep 2026

Kronos under the microscope

Kronos is a public foundation model for financial time series, built by its original authors. I tested the public Kronos-small checkpoint against simple baselines on CSI300, traced how its forecasts are formed, and added a bounded KV cache to its inference path.

My role
Evaluation, forecast tracing, inference optimization
Upstream
Kronos model and checkpoints by the Kronos authors
Status
Completed study of one public checkpoint

Limits: one checkpoint, one universe, one window. The Kronos result moved from +1.93% to +6.01% a year with the number of averaged forecast paths, and the ranking signal was not significant where tested (p = 0.295).

Read the case: question, role, evidence, limits →
eval/topk_backtest.py · CSI300 · Kronos-small
Cumulative excess return over CSI300 after costs: Kronos-small ends at plus 1.3 percent, reversal plus 0.7, random minus 2.6, momentum minus 16.3.
Bounded KV cache11.7× fasterBit-identical output · CPU, 400-bar lookback
Ranking signalp = 0.295Not significant (Newey–West, 5-path run)

Redrawn from the evaluation’s daily reports: compounded excess over CSI300 after costs. The annualized mean, +1.93%, is the figure quoted elsewhere.

03 · PhD research · in progress

Multi-wave mediation in CFPS

Does socioeconomic deprivation lead to depression through physical health, or do the two simply travel together? Using the China Family Panel Studies, I separate two questions that are easy to confuse: what would change under an intervention across people, and what happens within the same person over time.

Setting
Professor Dora Y. Zhang's research group, HKU
Data
China Family Panel Studies, multiple survey waves
Status
Analysis and methods write-up in progress

No effect estimates appear on this site until the analysis is final. Time ordering, identification assumptions and sensitivity checks come before any causal reading.

Research direction →
Methods note · multi-wave design
Deprivation, physical health and depression across four survey waves. Track A reads the exposure history into the mediator and outcome under a regime; Track B follows within-person lagged paths.
StatusIn progressNo estimates published
Two estimands
Track A · interventionalTrack B · within-person

Conceptual study design; no data and no estimates.

Earlier research.

Schematic illustration for generative models for microbiome data; not data.Duke University · 2024

Generative models for microbiome data

In Li Ma's research group, I studied generative models for microbiome data and evaluated generated samples through bioinformatics analyses on large datasets.

Collaborative research, May – Dec 2024.

Schematic illustration for optimization under uncertainty; not data.UC Berkeley

Optimization under uncertainty

In Ying Cui's research group, I worked on a semismooth Newton approach to risk-aware, nonconvex optimization and compared its computational behaviour with other solvers.

Collaborative research.

Schematic illustration for deep learning price forecasts; not data.Fudan University · 2023

Deep learning price forecasts

In Weiping Zhang's research group, we studied deep learning price forecasts and evaluated them through an A-share backtest with monthly portfolio rebalancing.

Group project, Jun – Sep 2023.

Schematic illustration for active topological glass; not data.Nankai University · 2022–23

Active topological glass

In Yao Li's research group, I used molecular dynamics simulation to study active topological glass under ellipsoidal confinement, and presented the work at the 21st National Condensed Matter Theory and Statistical Physics Conference.

Nov 2022 – Apr 2023.