The Minds Behind the Models: Zhenhao Sng

AlphaNova’s Competition 5 just concluded, and the numbers tell a story of serious rigour: more than 6,000 data scientists, quants, and engineers from around the world joined, yet fewer than 1,000 trading signals made it past our initial checks, and only 160 signals cleared the full gauntlet of overfitting, look‑ahead bias, and correlation filters. Behind every one of those surviving signals is a sharp, curious mind that found a way through.

In this series, we’re sitting down with some of those participants to hear about their backgrounds, their strategies, and the tools they actually use. These aren’t just quant profiles — they’re an honest look at how people from wildly different paths ended up competing in the same arena, and what they learned along the way.

First up: Zhenhao Sng

Zhenhao Sng: Building a Quant Toolkit with Ridge Regression and Rigorous Validation

Zhenhao is a third‑year engineering student at the Singapore University of Technology and Design, but he doesn’t fit neatly into one box. His degree bridges financial services, business analytics, operations research, psychology, and artificial intelligence — and he’s simultaneously preparing for the CFA Level I exam and gearing up for the WorldQuant BRAIN Research Programme. That interdisciplinary mix shows up in how he tackles quant problems: with the instinct of an engineer, the curiosity of a researcher, and the discipline of someone who’s learned the hard way that a shiny backtest doesn’t mean a signal works.

Like many participants, Zhenhao’s entry into quantitative finance was driven less by a single course and more by the pressure of competitions.

“What actually pulled me toward quant finance specifically was competing in things like the IMC Prosperity Challenge this year as well as the WorldQuant International Quant Championship. There’s something about a competition forcing you to ship a working model by a deadline that theory classes just don’t replicate. I’m still pretty early in this, though, so I’d love to have more opportunities to explore and challenge myself further!”

That attitude carried straight into his AlphaNova work. He didn’t come in with a grand, complex architecture. Instead, he focused on simple feature transformations — rolling normalisation, changes in cross‑sectional ranks, persistence effects, and state‑based features — and paired them with regularised linear models. The goal wasn’t to build the fanciest pipeline; it was to build something he could understand.

“My process was mainly to start with simple, causal feature transformations and test whether they captured a repeatable effect across assets and time. I generally used small, regularised linear models rather than more complex ones, as they made it easier to understand what was driving the signal and reduced the risk of fitting noise.”

In fact, when asked about the tool he reaches for most, he didn’t name a trendy deep‑learning framework. He named Ridge regression.

“It is simple, fast, and surprisingly effective when there are many related or noisy engineered features. I have also found myself paying much more attention to validation design. A model can look impressive on one split and completely disappear when tested across different periods or against shuffled labels, so robust validation has become just as important to me as the model itself.”

That rigorous validation mindset was crucial, because early on, many of his signals didn’t survive.

“Quite a few of my early models turned out to be overfit. Marc’s feedback helped me better understand how the platform’s overfitting check worked, which allowed me to improve my validation process and refine later iterations more carefully. The signals that survived tended to be relatively simple effects that remained stable across different periods and were sufficiently distinct from existing submissions.”

This is a theme we’ve seen across the competition: the signals that pass AlphaNova’s filters aren’t the ones with the most complex math — they’re the ones that prove themselves out‑of‑sample, period after period.

Zhenhao’s instinct for validation extends beyond quant finance. He recently competed in a Kaggle competition predicting car safety‑feature choices, where a subtle encoding error silently reconstructed the target variable — a data leakage trap that took days to catch. He also worked on a project optimising HIV funding allocation across African countries, using a mix of greedy optimisation, simulated annealing, and multi‑objective genetic algorithms (NSGA‑II). It’s a different domain from signal forecasting, but the same philosophy applies: build a model, then try hard to break it before you trust it.

“We built a model, then tried hard to break it before you trust it. That instinct is the same — whether it’s healthcare or finance.”

On the AI front, Zhenhao’s approach is pragmatic. He uses tools like ChatGPT and Claude as coding and research assistants, but he’s careful about where the decisions ultimately rest.

“I mainly use AI tools as a coding and research assistant. They are useful for brainstorming alternative feature ideas, debugging implementation issues, explaining unfamiliar techniques, and helping me structure experiments more systematically. But for modelling work especially, I still verify the code, check causality and data alignment, and test every idea against proper validation. The actual decision about whether something works still comes from the data itself.”

Connect with Zhenhao on LinkedIn and follow his growing journey in quant finance, engineering, and AI.

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 Zhenhao’s did.