Role: Product • Data • Full-Stack Engineering • Experimentation Design

Timeline: Summer 2026 — Present

Stack: Python, JavaScript, SQL, GitHub Actions, Alpaca Paper API, automated testing and research tooling

<aside> 🧭

I built a system that could prove when my trading ideas were wrong.

NewsTrader is a paper-only algorithmic trading research platform designed to separate useful evidence from convincing-looking noise.

</aside>

The problem

Algorithmic trading projects can look successful for the wrong reasons: incomplete data, look-ahead bias, idealized fills, tiny samples, or tuning against the same period used for evaluation.

I wanted the system to answer a harder question: does this strategy deserve trust at all?

What I built

• Separated historical simulation, prospective shadow observation, and paper execution so one type of evidence could not borrow credibility from another.
• Added data-quality gates, point-in-time validation, transaction-cost assumptions, and explicit review thresholds.
• Built operator tooling that shows health, freshness, missed runs, sample sizes, lifecycle states, and scenario outcomes before encouraging any performance interpretation.
• Added safety controls that keep live trading disabled and prevent automatic strategy promotion.

Research architecture

Historical research

Explore hypotheses and test whether they survive basic evidence and execution assumptions.

Prospective shadow

Observe future signals on data the strategy did not use during development.

Paper execution

Evaluate broker and operational behavior under separately controlled paper-only conditions.

What failed — and why that mattered

Several early ideas did not demonstrate a reliable edge. A broad support/resistance replay produced negative expectancy, small weekday samples did not justify hard-coded rules, and some hypotheses could not be evaluated honestly because required point-in-time features or reliable datasets were unavailable.

Instead of hiding those outcomes, I redesigned the product so failure and uncertainty became first-class results.

How the product changed

Measured system outcome