Models

Built to answer a question I couldn't answer by reading.

I'm not a quant, and none of this runs the portfolio. The models exist because judgment has blind spots, and some of those blind spots are measurable.

01 Why Build Anything

Every model here started the same way: a question that kept coming up in post-mortems, that I couldn't settle by thinking harder about it. Is this a pattern or a story? Does the thing I believe about this market actually show up in the data, or have I been rewarded for it by coincidence?

The useful output is rarely a signal. More often it's a correction to an intuition — finding out that an effect I was confident about was mostly sector composition, or that a screen I trusted stopped working in exactly the conditions where I most needed it.

A note on detail. What follows is deliberately calibrated: enough to show the work was genuinely specified, built, and run, and not enough to reconstruct anything proprietary. Where a model was developed in a professional capacity, the description is confined to method and general findings — no employer attribution, no live results, no production parameters.

02 Selected Work

Sector-neutral quality screen for Canadian small- and mid-caps

In production
The problem
Generic factor screens applied to the Canadian universe import enormous unintended sector bets. A naive value rank in this market is, most of the time, a levered bet on energy and materials — so the screen ends up expressing a commodity view I never intended to take, and its apparent efficacy is really just sector timing.
Approach
Rank within sector rather than across the universe, on a composite of profitability stability, accrual quality, and reinvestment rate. Winsorise inputs to handle the fat tails that resource names produce. Apply a liquidity floor before ranking rather than after, so the output is an actionable list rather than a theoretical one.
What it showed
Neutralising sector gave up a meaningful share of the raw spread — but the residual was far more stable across regimes, and the drawdown profile improved materially. Most of the headline performance of the un-neutralised version turned out to be one macro factor in disguise.
Limitations
Thin coverage in the smallest decile makes the fundamental inputs unreliable, and the screen degrades in sharp regime shifts where trailing fundamentals lag the turn. It generates a reading list; it does not generate a position.
Where it sits
Feeds stage two of the process — idea generation only. No output of this model reaches the portfolio without the full primary research stage.

[Model two — working title]

Research
The problem
What question kept recurring that you couldn't settle by reading? State it as a problem, not as a solution.
Approach
Universe, the core construction choice, and the one non-obvious design decision you made. Enough that a practitioner nods; not enough that they could rebuild it.
What it showed
The finding — stated directionally, including anything that contradicted what you expected going in. Contradictions are the most credible thing on a page like this.
Limitations
Where it breaks and what it cannot do.
Where it sits
Which stage of the process it informs, and what it is explicitly not allowed to decide.

[Model three — working title]

Retired
The problem
Consider making one of these a model that failed or was retired. A manager who shows one that didn't work is far more believable than one who shows three that did.
Approach
—
What it showed
—
Limitations
—
Where it sits
—
03 Discipline

How I try not to fool myself.

The failure mode in this work isn't a bug. It's building something that describes the past beautifully and tells you nothing about the future, then trusting it because it was expensive to build.

Next AI in Research →