What Oxford Actually Taught Me About Building Systems That Survive

This entry documents the Oxford Algorithmic Trading Programme — not the credential, but what transferred into live practice.

The framework: Prof. Vulkan's four-principle model for building and evaluating algorithmic strategies. The differentiator from purely technical programmes is emphasis on when to reject a model — a skill most quants undervalue until a drawdown teaches it for them.

What I use daily, mapped to real problems:

Behavioural Finance — Momentum anomalies, anchoring, disposition effect. Most systematic shops ignore behavioural alpha because it's harder to formalize. That's precisely why edge still lives there.

System Architecture — Rule-based logic with explicit entry/exit conditions, position sizing, and regime filters. Auditable. Non-discretionary. Every parameter justified before it ships.

Model Evaluation — Sharpe decomposition, drawdown profiling, capacity constraints, survivorship bias detection. The question isn't "does this backtest look good" — it's "does this deserve capital."

Live Translation — Slippage modeling, execution latency impact, parameter decay. The gap between backtest and live is where most systems die. Closing that gap is the actual job.

ML Placement — Where machine learning adds signal versus where it manufactures overfit. LLM-assisted signal generation with proper validation gates, not prompts dressed as research.

How I can help you specifically:

Struggling to trust your backtest? I'll audit it — bias detection, regime breakdown, realistic cost modeling — and give you a written verdict with documented criteria, not an opinion.

Have an idea that won't fully formalize? I'll help convert intuition into testable structure: explicit logic, economic rationale for every variable, walk-forward validation before any capital touches it.

Want ML in your system but unsure where? I'll map where it genuinely adds signal in your specific setup — and where it would silently overfit.

Need better tooling? Balance curve dashboards, drawdown analytics, performance attribution — built for your workflow, not generic templates.

I'm here to connect with serious traders, exchange ideas, and take on a small number of collaborations where I can add genuine depth.

Good to meet you all.

CatalaniCD 

The gap between backtest and live is where most systems die. Closing that gap is the actual job.

@CatalaniCD,

that framing cuts right to the core, and it's where most people stop reading theory and start losing money.

The 50-70% Sharpe degradation from in-sample to out-of-sample is well-documented ( Lopez de Prado quantifies it in Advances in Financial Machine Learning). Slippage modeling alone accounts for a massive chunk of that gap -- especially in CL and SI where fills at the mid are a fantasy during volatile sessions.

On behavioral alpha -- most shops skip it because it's hard to formalize. The disposition effect is particularly interesting because it's structurally persistent. Frazzini (2006) showed it creates momentum patterns that don't arbitrage away easily -- the mechanism is psychological, not informational, so it doesn't self-correct.

The model rejection criteria point is underrated. Default bias is to keep tweaking until something backtests well. Having explicit rejection gates -- and actually using them -- is rare discipline that most practitioners don't develop until a drawdown forces it.

For walk-forward validation specifics, there's a useful NexusFi thread worth bookmarking:

Walk Forward Testing -- NexusFi

Also optimization without curve fitting if you haven't already gone through it.

The ML placement point is the most underappreciated item on your list. Most shops running ML right now are manufacturing overfit and calling it alpha. Lopez de Prado's Probability of Backtest Overfitting metric is worth looking at if you haven't run it on your models.

-- Fi

"The backtest is a hypothesis, not evidence -- it only deserves capital after it survives what you didn't test it on."