The Minds Behind the Models: Utkarsh C Doshi

Utkarsh C. Doshi: The Researcher Who Makes His Process Deliberately Boring

Utkarsh C. Doshi has spent his career inside some of the most sophisticated quant operations in the world. He holds a dual degree from IIT Kharagpur — a B.Tech (Hons) in Metallurgical and Materials Engineering alongside an M.Tech in Financial Engineering. He's cleared CFA Level 1 and FRM Parts 1 and 2. He spent time in J.P. Morgan's Equity Derivatives Structuring Group, working on Delta One systematic trading strategies and factor investing, and he's now a Senior Researcher at WorldQuant.

You'd expect someone with that résumé to talk about exotic models and elaborate architectures. Instead, his answers read like a quiet argument for restraint.

"I start from one clean, economically-sensible idea rather than a kitchen sink of features, and then most of the work goes into making it robust — controlling risk exposures, favoring consistency over a big one-off number, and keeping trading costs in check."

That's the whole philosophy, stated plainly. And it runs counter to the instinct that dominates a lot of quantitative competition entries.

One Idea, Pushed as Far as It Goes

When asked about a recent project he was excited about, Utkarsh pointed to the AlphaNova competitions themselves — but for a specific reason.

"In the last one I got a lot out of taking a single, simple idea and pushing it as far as it would go — spending most of my time making it more robust rather than piling on complexity. A good reminder that in noisy markets, cleaning up a signal often matters more than finding a new one."

That last sentence is worth sitting with. Most participants in a signal forecasting competition spend their time searching for new features, new interactions, new model classes. Utkarsh spent his making a single signal more durable. In a field where overfitting is the primary failure mode, that's not a lack of ambition — it's a deliberate strategy.

It also echoes a pattern across this series. Ismam retired an entire architecture rather than tune it. Mathurin built an evaluator specifically to catch himself overfitting. Alok used Ridge regression on cross‑sectional ranks and let the validation layer do the work. None of them won on model sophistication. All of them won on discipline.

Trying to Break His Own Ideas

Asked about his process, Utkarsh described something that sounds almost adversarial toward himself.

"Fairly disciplined, and a bit boring on purpose. I start from one clean, economically-sensible idea rather than a kitchen sink of features, and then most of the work goes into making it robust — controlling risk exposures, favoring consistency over a big one-off number, and keeping trading costs in check. I also spend a lot of effort trying to break my own ideas, because a surprising amount of apparent edge turns out to be overfitting or subtle look-ahead."

Trying to break my own ideas. That's the same instinct Mathurin described when he built a surrogate model of the validation‑to‑test haircut — treating the gap between local and server scores as the actual object of study rather than an inconvenience. It's the instinct that separates researchers who last from researchers who get lucky once.

The phrase boring on purpose is also doing real work. There's a version of quant research that's exciting — big models, novel architectures, dramatic backtests. And there's a version that's tedious and careful and quietly more profitable. Utkarsh has chosen the second.

Simple Models, Fast Tooling

On the tools question, he stayed consistent with everything else he'd said.

"Lately I keep coming back to simple, well-regularized models and fast, clean data tooling over anything fancy. Being able to iterate quickly on large datasets, plus techniques that reduce noise rather than add parameters — that combination has done more for my results than any single complex model."

Techniques that reduce noise rather than add parameters. It's a compact statement of a principle that shows up across the entire series. Kushpreet reached for distributional forecasts rather than point estimates because fat tails demanded it. Zhenhao named Ridge regression as his most-used tool. Alok calibrated his evaluator to the server at roughly 0.52x rather than chase a higher local number.

In each case, the winning move was subtraction, not addition.

AI as Sanity Check, Not Source of Alpha

Utkarsh uses AI heavily — but his framing is notably careful.

"I lean on an AI coding assistant for the heavy lifting of my research loop — writing and running experiments, organizing results, keeping notes on what worked and what didn't — so I can spend more time on ideas than on wiring up code. It's also a useful sanity check when a result looks too good to be true. It won't hand you alpha, but it makes the whole cycle much faster."

That last line is a useful corrective to some of the more enthusiastic takes in this series. Taro credited AI with enabling results he couldn't otherwise have achieved. Saksham used it to generate 28 candidate signals in two weeks. Utkarsh's version is more measured: AI accelerates the loop, but it doesn't produce the idea, and it doesn't decide what's real.

It won't hand you alpha. For a researcher at WorldQuant, that's probably the most important thing to keep in mind.

What We Can All Learn

Utkarsh's profile is short and understated, but it contains three lessons that apply broadly:

One clean idea, pushed to robustness, beats a kitchen sink of features. In noisy markets, cleaning up a signal often matters more than finding a new one.
Spend effort trying to break your own ideas. A surprising amount of apparent edge is overfitting or subtle look‑ahead. The researchers who last are the ones who go looking for it.
Simple, well‑regularized models plus fast tooling outperform complexity. The combination of quick iteration and noise reduction has done more for his results than any single complex model.

That's a philosophy built on subtraction rather than addition — and it's 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 Utkarsh's did.