A research system
with an audit trail.
Helix connects signal research, backtesting, portfolio controls, validation and paper trading, so each step can be inspected before a result reaches a decision. Explore the public web front ↗
- My role
- Design, implementation and validation
- Builds on
- Qlib (one ML sleeve); HKUDS Vibe-Trading factor engine (MIT)
- Since
- June 2026
- Markets
- US, China A-shares, Hong Kong
- Code
- Private repository
From one neuron to a trading book. The prices are real, from Helix’s data lake; the equity curve is illustrative, and Helix’s real combined book fails the gate it runs into (see Evidence).
Read the script
- One neuron. Inputs add up; past a threshold, it fires: y = σ(Σ wᵢxᵢ + b).
- Stack them: deep learning. Helix’s torch MLP, 64 → 32 hidden units. Forward to predict, backward to learn.
- Feed it numbers. 9,768,394 real daily bars from the lake, 2015 to 2026.
- Numbers become markets. 4,141 stocks: US 1,837 · China A 1,801 · Hong Kong 503.
- Rank them. Build a strategy. One of 576 factors, ranked: long the top fifth, short the bottom, where the market allows.
- Compound the edge. Net of every cost, then a deflated-Sharpe gate that counts every trial ever run. The curve is illustrative and labelled so.
- Then trade. Paper first, through IBKR, in each market’s own session. Live trading stays fail-closed.
- From one neuron to a trading book. Data → signal → backtest → combine → validate → paper.
Can a signal stay traceable from research to paper execution?
A signal should carry its data provenance, a consistent backtest reading and risk checks before it reaches a paper workflow.
Built end to end, on others’ foundations.
I designed and built the shared signal core, factor lab, backtest engines, a risk-limited portfolio allocator, validation gates and the simulated paper-trading loop. Qlib powers one ML sleeve; the factor engine and factor library are vendored from HKUDS Vibe-Trading under the MIT licence.
Checks along the chain.
Signal parity checks, embargoed walk-forward research, cost-aware backtests and run records keyed by code, configuration and data. The map shows those components; it is not a performance chart.


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.
Paper evidence has a boundary.
- Point-in-time universe masking is not routed through strategy backtests, so absolute returns are overstated by about five points a year; relative comparisons hold.
- Delisted price coverage is incomplete.
- The recorded paper history includes replayed days.
- Real-capital execution remains gated.