The Minds Behind the Models: Ismam Tabriz
Ismam Tabriz: The Self‑Taught Engineer Who Treats Every Backtest with Suspicion
Ismam Tabriz doesn’t fit the typical quant mould. His formal degrees are in psychology, not computer science or finance. He grew up in Khulna, Bangladesh, and started programming around age nine — long before any university accepted him into an engineering programme. Today, he works as a Senior Engineer at Artixcore, building scalable backend systems and AI‑powered infrastructure. But it’s his instinct for rigorous, adversarial self‑testing that made him stand out in AlphaNova’s Competition 5.
“I studied social work and psychology because I needed to pursue a formal undergraduate education, but programming has been part of my life for much longer. I started programming around 2009, when I was about nine years old. I was mostly self‑taught, driven by curiosity and the simple excitement of making a computer do something that previously existed only in my head.”
That curiosity eventually led him to markets, and then to a question that sits at the core of quantitative finance: can we systematically find useful patterns in noisy financial data without fooling ourselves?
“What I enjoy about quant research is that an idea cannot survive just because it sounds clever. You have to encode it, test it on unseen data, account for transaction costs, challenge your assumptions, and see whether the signal survives.”
AlphaNova’s obfuscated data made the problem even more interesting for him. Without the ability to lean on a story about what a feature “means,” he had to work from statistical structure alone.
“AlphaNova makes that even more interesting because the financial data is obfuscated. You cannot depend too heavily on a story about what you think a feature represents. You have to work from the statistical structure itself and pay attention to causality, robustness, and generalization. For someone who learned programming mainly by building and experimenting, that feels very natural.”
Treating Research Like a Small Engineering Project
Rather than treating AlphaNova as a one‑off competition submission, Ismam approached it as a research engineering project. He built a repeatable workflow around feature experimentation, causal validation, turnover‑aware evaluation, deterministic testing, and signal robustness.
“I have been treating it less like a single competition submission and more like a small research engineering project. I built a repeatable workflow around feature experimentation, causal validation, turnover‑aware evaluation, deterministic testing, and signal robustness.”
When an early architecture didn’t perform well enough, he didn’t endlessly tune it to make a prettier backtest. He retired it and started again from first principles.
“One of my earlier approaches did not perform well enough, and instead of endlessly tuning it until I got a prettier backtest, I retired the architecture and started again from first principles. I like that part of engineering. A failed experiment is not wasted if it tells you clearly what not to build next.”
That discipline carried into his modelling choices. He became more interested in the opposite of complexity: how little complexity can I get away with while still extracting something useful?
“Probably regularization combined with very strict validation. It is tempting, especially with modern machine learning, to keep increasing model complexity whenever performance is disappointing. I have become more interested in the opposite question: how little complexity can I get away with while still extracting something useful?”
His go‑to tools reflect that philosophy: Ridge regression, shallow tree models, and a battery of validation checks — walk‑forward evaluation, shuffled null experiments, causality tests, and transaction‑cost‑aware scoring.
“I have become quite skeptical of beautiful backtests. If the result disappears after costs, changes dramatically under a small perturbation, or contains subtle future information, then the beauty was mostly an illusion.”
Building Signals That Survive Turnover and Causality
In the AlphaNova competition, Ismam’s strategy was deliberately conservative. Instead of constructing dozens of rolling time‑series indicators, he focused on the cross‑section at each timestamp. His feature representations included cross‑sectional ranks, differences between features, compact nonlinear interactions, disagreement measures, and consensus statistics. He then blended low‑capacity models — strongly regularized Ridge models plus a shallow boosted‑tree component.
Turnover was a major consideration.
“If transaction costs are part of the evaluation, maximizing predictive accuracy without thinking about how frequently the portfolio changes is incomplete. A signal can be statistically interesting and still be economically useless.”
He used causal smoothing at the portfolio‑output level to reduce unnecessary turnover, while ensuring future information couldn’t affect earlier predictions. He also controlled for persistent instrument‑level bias, handled extreme predictions robustly, and maintained the required cross‑sectional de‑meaning.
“I would rather have a modest signal that survives unseen data, transaction costs, causality tests, and perturbations than an extraordinary‑looking backtest held together by one fragile assumption.”
He’s also increasingly interested in signal novelty — producing a genuinely different source of predictive structure rather than a slightly modified version of an existing strategy.
AI as a Challenger, Not an Oracle
Ismam uses AI heavily, but not as a black‑box oracle. Instead, he treats LLMs and coding agents as collaborators that can implement, review, debug, test, document, and challenge his ideas. He even works with multiple specialised agents — each taking a different role such as engineering, reviewing, research, security, or monitoring.
“Very heavily, but I do not treat AI as an oracle. I use LLMs and coding agents for research brainstorming, implementation, code review, debugging, test generation, documentation, infrastructure work, and challenging my own ideas.”
The shift in how he interacts with AI has changed his thinking. Rather than asking “Can you write this code?”, he now asks much sharper questions:
“Where could this experiment be leaking information? What assumption am I making without realizing it? What would cause this strategy to fail? How could this feature accidentally encode identity? What experiment would falsify my hypothesis?”
For him, AI dramatically shortens the distance between an idea and a working experiment. But the final verdict still belongs to reality.
“AI can dramatically shorten the distance between an idea and a working experiment. That means one person can investigate far more possibilities than before. But I still want reality to have the final vote. For quantitative research, that means reproducible experiments, unseen data, causality checks, transaction costs, and actual out‑of‑sample results.”
A Lifelong Builder’s Mindset
What stands out most about Ismam is not any single technique, but the way he approaches the entire process. He is self‑taught, endlessly curious, and deeply skeptical of anything that looks too good. His path into quant finance wasn’t through a traditional pipeline — it came from years of building, breaking, and studying what went wrong.
“I am ultimately a fairly simple Bangladeshi Muslim who has been fascinated by programming since childhood. I did not come through a traditional engineering or quantitative‑finance path. Most of what I know technically came from years of curiosity, building things, breaking things, studying what went wrong, and trying again. AI has simply made that lifelong process much faster.”
That mindset — humble, rigorous, and relentlessly experimental — is exactly what the AlphaNova community is built on.
Stay tuned for more profiles from the AlphaNova community. If you’d like to be featured, reach out—we’d love to share your story.
Join the latest AlphaNova competition and see if your signals can survive the same rigorous tests that Ismam’s did.