Helix
A research and paper-trading system for US, China A-share and Hong Kong equities, with a validation gate that can reject its own strategies.
Read the Helix case
Statistics PhD researcher · The University of Hong Kong
Uncertainty, drawn to scale.
I am Liyu Cao. I study causal inference and statistical machine learning, and build quantitative research systems that make assumptions, uncertainty and failures visible.
Each line is one possible future. The rings hold 95% of them.
Research training across five universities
The University of Hong KongDuke UniversityUC BerkeleyFudan UniversityNankai University
Research questions
What can the data actually tell us?
Longitudinal data, mediation, time ordering and the assumptions a causal interpretation needs. In CFPS I keep between-person and within-person questions apart.
Research direction →Generative models, the evaluation of synthetic data and uncertainty in high-dimensional settings, starting from microbiome data at Duke.
Research direction →Point-in-time data, out-of-sample testing, trading costs and the path from a forecast to a portfolio.
Research direction →
CFPS · study design, conceptual
Four checks before a result counts.
Read the notes →Choose the estimand.
A model is only useful once the quantity and the decision are clear.
Respect time.
Use only the information that existed at the point of inference or decision.
Make comparison fair.
A simple alternative makes a complex result earn its place.
Try to break it.
Sensitivity checks show which assumptions carry the conclusion.
Selected work.
All projects →Kronos under the microscope
A public financial foundation model tested against simple baselines and made faster without changing its output.
Read the Kronos case
Multi-wave mediation in CFPS
Deprivation, physical health and depression across survey waves, with between- and within-person questions kept apart.
See the research directionWhat the evidence supports, and where it stops.
Kronos-small against simple baselines
Public checkpoint · CSI300, 225 trading dates, Jul 2024 – Jun 2025 · annualized excess return after costs, 10 averaged forecast paths
+1.93%vs +1.66% for the reversal baseline. A 5-path run gave +6.01%, and its ranking signal was not significant (p = 0.295).
Kronos inference with a bounded KV cache
One A-share 5-minute series · CPU · 400-bar lookback, 100 greedy steps
11.7×faster, with output bit-identical to the original path; exact only while the context fits the model's window.
Survivorship in a Helix backtest universe
Equal-weight S&P 500 universe from 2015: today's members throughout vs each stock held only while it was a member
−5.08 ptsa year: 16.14% falls to 11.06%. Why the Helix pages quote no absolute backtest returns.
Multi-wave mediation in CFPS
China Family Panel Studies · between-person and within-person designs
In progressNo effect estimates are published here until the analysis is final.
Each figure comes from a specific run in my own project files, with its setting stated beside it. None of them is live trading performance.
Research is better when it can be questioned.
For questions about statistics, causal methods or quantitative research, I would be glad to hear from you.