01 Home02 Research03 Projects04 Notes05 About06 Lab
HKU Statistics PhD researcher

Uncertainty World, Infinite Progress

I am Liyu Cao. I study causal inference and statistical machine learning, I build quantitative research systems that make assumptions, uncertainty and failures visible.

Cross-section t = 0.60
A acrossB up

Each line is one possible future. The rings hold 95% of them.

Chance, at every scale.

Sixty seconds · painted in ink
01 · Quantum uncertainty

At the smallest scale, nature rolls dice.

Each dot is one measurement of where an electron is. No single answer can be predicted; together they draw the exact shape of the odds.

Read the script
  1. Quantum uncertaintyAt the smallest scale, nature rolls dice. Each dot is one measurement of where an electron is. No single answer can be predicted; together they draw the exact shape of the odds.
  2. Brownian motionA pollen grain in water never sits still. Water molecules kick it from every side at random. Einstein explained the jitter in 1905; measuring it proved atoms are real.
  3. DiffusionDrop ink in water: countless random paths paint one smooth cloud. No single molecule is predictable, yet the cloud spreads like clockwork: four times the time, twice the spread.
  4. The Poisson processRain falls at random. A shower is predictable. Count the drops each second: the count jumps around, but its shape is Poisson’s law, the same law that counts calls to a help desk.
  5. Random walksPrices wander like pollen. In 1900, five years before Einstein, Louis Bachelier used the same random walk for the Paris stock market. Tomorrow’s price is unknown; its range is not.
  6. ProbabilityOne electron, one raindrop, one trade: unpredictable. Millions: a law. Probability finds the shape inside the randomness. Uncertainty, drawn to scale.
Play with probability in the Lab →
Research questions

What can the data actually tell us?

Mediation over several survey waves, where time ordering decides which effects are identifiable.

Judging synthetic data by what downstream analyses still find in it, starting from microbiome data at Duke.

Point-in-time data, costs and simple baselines between a forecast and a portfolio.

CFPS study design over four survey waves, conceptual. Schematic: observed and generated microbiome compositions; not data. The Helix research terminal, a saved snapshot of an in-sample backtest. CFPS · study design, conceptual

Four checks before a result counts.

Read the notes →
01 · Question

Choose the estimand.

A model is only useful once the quantity and the decision are clear.

In practice: a mediation timeline
02 · Data

Respect time.

Use only the information that existed at the point of inference or decision.

In practice: survivorship
03 · Baseline

Make comparison fair.

A simple alternative makes a complex result earn its place.

In practice: forecasts vs trades
04 · Stress test

Try to break it.

Sensitivity checks show which assumptions carry the conclusion.

In practice: a control that can fail

Selected work.

All projects →
Independent system · 2026

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
The Helix research terminal: signal-parity and risk checks pass, deflated-Sharpe validation fails.
Independent evaluation · 2026

Kronos under the microscope

A public financial foundation model tested against simple baselines and made faster without changing its output.

Read the Kronos case
Cumulative excess return over CSI300 for Kronos-small and three baselines.
PhD research · in progress

Multi-wave mediation in CFPS

Deprivation, physical health and depression across survey waves, with between- and within-person questions kept apart.

See the research direction
CFPS study design over four survey waves, conceptual.

What the evidence supports, and where it stops.

Kronos-small · CSI300 · Jul 2024 – Jun 2025 · after costs

A narrow edge over the simple.

Kronos-small+1.93%Reversal+1.66%Random−2.79%Momentum−18.80%5 paths−20%−10%0

One checkpoint, one window. A 5-path run showed a larger edge, and its ranking signal was not significant (p = 0.295).

Kronos case →
Kronos inference · CPU · one A-share series

11.7× faster, the same answer.

Time for 100 stepsOriginal3.80 sCached0.32 s · 11.7×Context window (bars)400 lookback+100512 maxfalls back

Exact only while lookback and forecast fit the 512-bar context; beyond that it falls back to the original path.

Note: exact, then fast →
Helix · equal-weight S&P 500 universe from 2015

Levels fall. Comparisons hold.

Today’s membersMembers at the time16.14%11.06%equal-weight+0.80%+1.59%momentum vs EW−5.08 pts a year

Absolute returns read about 5 points a year too high; active comparisons stay sound. So Helix quotes no absolute returns.

Note: survivors →

Each figure is drawn from a specific run in my own project files. None of them is live trading performance.

Research training across five universities

The University of Hong KongDuke UniversityUC BerkeleyFudan UniversityNankai University

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.

Send an email →