A market-neutral fund that stays level when AI moves the market

Finalist · McGill-FIAM Asset Management Hackathon 2026

The fund

AI is starting to move markets, and machines tend to think alike. QUARCC Purple is built to stay level when they do. It's a $1 billion fund that buys the strongest US companies and shorts the weakest with the same money on both sides, so the market's swings cancel out. In 2022, when the S&P 500 lost 18%, it made 1.3%. Machine learning ranks the stocks, code enforces every rule and eight AI agents read, explain and challenge the book but never trade.

Results

Less return than the market, with a fraction of the risk. Backtest, January 2021 to August 2026, after all costs.

	QUARCC Purple	S&P 500
Max drawdown	−2.99%	−25.4%
Volatility	3.96%	15.1%
Beta to the S&P 500	0.012	1.00
Ulcer index	0.52%	8.10%
Sharpe ratio	0.91	0.70
Sortino ratio	4.21	1.51
Return a year	7.0%	13.5%
$1bn became	$1.467bn	$2.046bn

How it works

Every month a machine-learning model scores every US stock on about 150 signals. An optimiser turns those scores into a balanced book of about 300 longs and shorts, traded slowly over five days, and code checks every rule on every day.

The AI agents

All eight run on a local model on our own machines and never place a trade. Their records are in finals/ai/.

Agent	What it does
Rules reader	Reads the competition rules and checks them against the fund's limits
Triage	Sorts our companies' news filings into good, bad or neutral
Analyst	Reads the latest reports of our ten biggest longs and shorts
Explainer	Explains in plain words why we held our biggest winners and losers
Memo writer	Writes the investor letter, with every number checked by code
Devil's advocate	Attacks the portfolio and lets the data answer
Risk desk	Writes the risk note on the final book, every level set by code
AI crowd	Sixty AI investors pick stocks, to measure how crowded our book is

Check it yourself

Install uv once. These need no data.

To	Run
Check every rule on every day and recompute the headline from the locked results	uv run run.py
Run the tests	uv run pytest
Rebuild the deck, byte for byte	uv run tools/deck/build.py

The fund itself reads licensed data from WRDS and the organisers that cannot be shared. With that data in place, these rebuild everything.

To	Run
Prove the code, inputs and 659 outputs match the lock	uv run fiam verify
Rebuild the whole fund from the raw data, then compare	uv run fiam rerun then uv run fiam compare
Rebuild the website's numbers	uv run python -m tools.pack then uv run python tools/build_site.py
Write the whole repo as one file, MAIN.py	uv run python tools/build_main.py

A rerun on an Apple Silicon Mac matches the lock byte for byte. The deck build needs Node and, once, pnpm install and pnpm exec playwright-core install chromium in site/. On Windows, set PYTHONUTF8=1 first.

What is where

Folder	Holds
finals/	The fund facts, the deck, every figure they show and the AI agents' records
fiam/	The fund in fiam/vfinal/, the AI agents in fiam/agents/ and the shared stock and file helpers
frozen/	The lock and the fund's locked results, the mandate templates and the agents' frozen inputs
tools/	The deck builder, the numbers pack, the site data builder, the data tools and the one-file builder
site/	The website and the film
data/	Small public series from FRED and Kenneth French
tests/	The tests

Team

Name	Role	Studies	Links
Nicholas Xydis	AI & Research	Concordia, Computer Science	LinkedIn · GitHub
Yohannah Ilieva	Risk Management	Concordia, Mathematical & Computational Finance	LinkedIn
Andrew Koumoudjian	Portfolio Construction	Concordia, Finance	LinkedIn
Nicolas Romero	Data & Finance	Concordia, Finance, Mathematics & Statistics	LinkedIn
Nokto Lapointe	Machine Learning	McGill, Honours Physics & Computer Science

Credits and license

Data · FRED, the Kenneth French library, Arcana and Carbon Arc (totals only). The S&P 500 series belongs to S&P Dow Jones Indices.

License · The code is under the MIT License. The data files keep their own terms.