AI FX - A Small Experiment in a Big Game
Institutional Trading Systems
I’ve spent some time around trading systems.
At Merrill Lynch, I supported the Order Audit Trail System (OATS), a compliance system required by NASDAQ. Its purpose is simple in theory but critical in practice: ensure institutions aren’t exploiting client activity for unfair advantage. Systems like that are designed to surface patterns that shouldn’t exist.
At CIBC, I led a team supporting the bank’s automated (no-touch) FOREX trading system. This was a mission-critical platform where downtime directly impacted revenue. According to the runbook, the system generated roughly $4M CAD per hour, and up to $40M during peak market activity.
That experience planted a seed.
I spent some time researching automated FOREX trading and came to a fairly blunt conclusion: this is a game that heavily favors institutions.
Not because the strategies are secret, but because of scale and positioning.
- Trades are executed in large volumes, where small statistical edges compound
- Systems are physically colocated near exchange infrastructure, reducing latency
A retail trader, working with limited capital and higher latency, is playing a very different game.
Moving on for a While
So I parked the idea.
Instead, I focused on other projects, including AI Hydra, a reinforcement learning sandbox for experimenting with live AI systems and hyperparameters.
The Lightbulb
Recently, I came across an extensive YouTube series: Forex bot & backtest system with Python for beginners.
It walks through building a basic trading bot using rule-based strategies.
And that triggered a thought:
- What if those rules were replaced or augmented with a neural network?
I’m not going into this expecting to “crack” FOREX trading.
But I am confident in two things:
- I’ll learn a lot
- There’s room to experiment with fast, iterative approaches that only act on high-confidence opportunities
FOREX also offers something compelling: a massive pool of historical data that can be treated as training material.
Planting the Seed
So I created a repo: GitHub Repository. I set up a basic framework, added a GitHub Action for publishing, and pushed an initial package to PyPI.
It’s early. Very early. I’ve seen how the big machine works… now I’m building a smaller one to understand it.
The seed is in the ground.
Now it’s a question of what grows.