OLIVER SCOTT
Open to research projects, mentoring and future quant internships

Oliver Scott

Secondary school student · planning a Mathematics / Computer Science degree
Expected university graduation 2031 · Brisbane

I'm a Brisbane school student who builds and tests quantitative ideas end to end: derivatives pricing, market making, matching engines, and an event-driven study of ASX junior miners now running as a live paper test.

Deep research

Research note

One question followed all the way: hypothesis, data, validation, execution, and a live forward test.

Live paper test since Nov 2026

Do ASX junior miners keep moving after good news?

An event-study framework over ASX resource stocks: a 72-cell search over buy delay and hold length, chosen in-sample and judged out-of-sample, placebo tests, limit-order and stop-loss tests, and a cash-runway filter read from quarterly Appendix 5B reports.

event studyout-of-sampleplacebo testslive systemcode on request

Read the note →

Heatmap of net return by buy delay and hold length
Breadth

Projects

Each has tested code, a reproducible figure script and a short write-up.

Research · Derivatives

Options pricing and implied volatility

Black-Scholes, a CRR lattice with early exercise, Monte Carlo with variance reduction, analytic Greeks and a robust implied-vol solver.

PythonNumPySciPypytest
4.2×variance reduction from a control variate
Trading · Microstructure

Market making under inventory risk

Avellaneda-Stoikov quoting against a symmetric benchmark: how skewing quotes by inventory trades a little mean P&L for half the risk.

stochastic controlPoisson fillssimulation
2.0×mean/std of P&L vs naive quoting
Development · Systems

Limit order book matching engine

Price-time priority with limit, market and cancel messages, property-style tests, and a one-million-message latency benchmark.

data structuresbenchmarkingtesting
450kmessages/s, p99 6.06 µs (pure Python)
Trading · Risk

Bet sizing: Kelly and estimation error

Growth versus fraction bet, two kinds of drawdown, and why full Kelly on an estimated edge can earn nothing.

probabilityMonte Carlorisk
0% vs 57%growth: full vs half Kelly on a noisy edge
Research · Statistical arbitrage

Pairs trading on ASX 200 constituents

Cointegration screening, Ornstein-Uhlenbeck half-lives, walk-forward selection and borrow costs.

Planned
Research · Volatility

Forecasting realised volatility

GARCH(1,1) vs HAR-RV vs a simple EWMA, scored with QLIKE on out-of-sample data.

Planned
Development · Performance

C++ port of the matching engine

Intrusive lists and a price-indexed array, targeting over 10M messages/s, benchmarked against the Python version.

Planned
Depth

More than the projects

Where I'm heading next, how I practise, and the maths running live.

Toolkit

Skills, with evidence

Each skill links to where it's used. Levels are honest: proficient, working knowledge, learning.

Mathematics

Programming

  • Python (NumPy, pandas, SciPy)proficient: every project
  • Testing (pytest)proficient: 33 tests across four public repos
  • Performanceworking: latency benchmark, implied vol 2.7 ms → 81 µs
  • C++learning: matching-engine port in progress
  • JavaScriptworking: demos · Git, Linux, SQL basics

Markets

  • Options & volatilityworking: pricing, Greeks, implied vol
  • Microstructureworking: order books, market making, auctions
  • Event studies & backtestingproficient: out-of-sample design, costs, biases
  • ASX mechanicsworking: announcements, Appendix 5B, halts, auctions

How I test an idea

No look-ahead

Signals use only data known at that close; trades enter the next day.

Out of sample

Choose on one period, judge on a later one, unchanged.

Placebos

Random dates for the same stocks. If they do as well, it's noise.

Real costs

Spread and impact scaled to liquidity; size capped by volume.

Theory checks

Simulations must match closed forms where they exist.

Paper first

Live forward test before any capital.

Transparency

How I built this

I use AI coding assistants the way I'd use a strong colleague: to draft code, suggest tests and review my reasoning. Every result here was run, checked against theory or an independent method, and is reproducible from a script in the repo.

Where I found and fixed problems, the write-ups say so: the implied-vol edge cases, the Kelly drawdown mix-up, the ranking rule that drifted to 120-day holds.

  1. Question first. I write the hypothesis and what would disprove it before any code.
  2. Test against truth. Closed forms, textbook values, planted effects and shuffled data.
  3. Rebuild to understand. I re-implement core pieces from a blank file and extend each project in my own direction.
  4. Show the limits. Every page lists what could make the result wrong.