Skip to content
View thirdbrew's full-sized avatar

Block or report thirdbrew

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
thirdbrew/README.md

Hey, I'm Charlie

I'm an Industrial & Systems Engineering student at Georgia Tech. Most of what I build sits where operations research meets money: optimization, simulation, and statistics pointed at decisions that cost something if you get them wrong.

The habit that shows up in all of it: I decide what "working" means before I run the test, then report the result against that bar. Sometimes the answer is that my idea doesn't work. Those results are in my repos too, and they're the ones I learned the most from.

This summer I was a Global Innovation Intern at Americold, building simulation and slotting tools for cold-chain warehouse throughput. Outside of that I run a systematic trading system I've been building and breaking for a while. The research I can publish is below, with the warehouse work on synthetic data.


What I work with

Core   Python pandas NumPy SQL Jupyter

Modeling   SciPy statsmodels LP / MIP Discrete-Event Sim

Engineering   Git pytest mypy


Selected work

Warehouse Layer-Pick Slotting Engine  ·  Python discrete-event sim

Which pallet should occupy which pick face, right now, and what moves next. A rules engine that emits advisory ADD/REMOVE/WAIT calls against a live floor state, plus a simulator that replays a full pick day against them. 84 modules, 295 tests, runs end to end on a seeded synthetic site. The customer feed it was developed against is not in the repository in any form, at any point in its history. On seed 7 its default policy takes the floor from a 77% dead-slot rate to 12%, and the waved layer-pick rate from 26.3% to 40.0%, weighted by the layers actually waved rather than averaged across hours. The opt-in resupply-rack policy reaches 10% and 48.3%.

Every rule is a pure function over one snapshot, which is what makes attribution possible: drop one rule, re-run, read what it was worth. That is how I caught the real problem. The first version I published moved dead slots 40pp and the layer-pick rate by exactly zero, every hour of the simulated day. The rule that fills a freed face was disabled by default and gated a second time, so the engine could evict and nothing else. The number I had led with was the one that couldn't tell.

Systematic Trading Research  ·  Python pandas SciPy

A 70/30 equity + managed-futures allocation, a swing sleeve, and the statistical machinery that decides whether either is worth funding. The pre-registration harness refuses to grade a hypothesis until its pass/fail bar is committed and pushed, because a bar still sitting in your working tree can be amended once you've seen the result. Trials are counted and the significance test is deflated for them.

Eight hypotheses registered, five graded, four missed their bars. The fifth prints FAIL but confirms its hypothesis, because its criteria were written the wrong way round. All five are published with their registrations and verdicts intact. The one I'd point at first misses its own bar by 0.0057 Sharpe and then argues, against my interest, that the criterion it nearly passed is the least robust of the three.

Cointegration Pairs Study  ·  Python pandas NumPy

Does pairs trading survive honest statistics? On a point-in-time S&P 500 universe (829 tickers, 2011–2026), a naive cointegration screen calls 173,594 pairs significant where chance alone predicts 122,920. After Benjamini–Hochberg FDR control 1,446 survive, and traded out of sample they return Sharpe −0.25 against a pre-registered bar of 0.50, and still −0.19 at zero cost.

Each of the three hypotheses was frozen and pushed to a remote before the run that grades it. An adversarial review then reversed one of the published conclusions: the trading path normalised its z-score with a statistic computed from the future. The reversal is documented, not patched.

Limit Order Book & Market Making  ·  Python pytest

A price-time-priority matching engine, a synthetic market with informed traders in it, and two market makers compared on a risk/return frontier. The naive fixed spread dominates the Avellaneda–Stoikov closed form at 7 of 7 risk levels, and the P&L decomposition says why: adverse selection. A-S prices inventory risk, and its derivation contains no informed traders. 56 tests.

Realistic Backtesting Engine  ·  Python pandas NumPy

A bar-by-bar equity backtester built so lookahead bias is impossible by construction rather than by remembering to avoid it. One place in the engine turns a decision into a fill, and it can only act on the previous bar's target. Every fill pays half-spread, slippage, and commission, all applied against the trader, which is the honest direction.

I used it on a naive EMA(5/8) crossover. It returned +6.7% over five years against +134% for buy-and-hold, with a worse drawdown and a mean daily return you can't distinguish from zero. That's the result, and documenting it cleanly was the point.


Performance figures are backtest output over a stated window, net of stated costs. The assumptions are written down, and I'm happy to walk through them.

charlesbmay3@gmail.com

Pinned Loading

  1. Back-Testing-Engine Back-Testing-Engine Public

    Bar-by-bar equity backtester with structural lookahead-bias prevention and realistic cost modeling.

    Jupyter Notebook