The Minds Behind the Models: Vibhav Agrawal

Vibhav Agrawal: The Physics Student Who Borrowed Feynman's Path Integrals

Vibhav Agrawal is still an undergraduate. He's studying for an MSci in Physics with Theoretical Physics at Imperial College London, expecting First Class honours in 2028. But he's already a member of the Algorithmic Trading Society, has been selected onto a quant trading team at Queen's Tower Capital — Imperial's student‑run fund — and produced a Competition 5 entry with one of the more unusual conceptual origins we've seen.

His route into quant finance started with a question about transfer.

"I've always had a deep fascination with statistics and data, it's one of those few things I really believe is 'true' in an absolute sense. So, naturally, I found myself at quant finance. I came across the field when I was exploring how I can apply physics outside physics. My personal tutor pointed me to research papers within his department on physics in options pricing and trading, which was when I became intrigued by how the tools transfer."

That word — transfer — is the theme of his whole approach. Physics has spent centuries building mathematical machinery for systems that resist intuition. Quantitative finance faces the same problem. The question is whether the tools carry across.

From Path Integrals to Signal Ensembles

Vibhav's AlphaNova strategy started with the standard machine learning approaches — he wanted to use them himself rather than just read about them. But the interesting part came from somewhere else entirely: a Quantum Physics lecture on Feynman path integrals.

"This is where instead of asking which single path a particle takes, you assign a probability amplitude to every possible path and sum over all of them. When I learnt this, I immediately thought about how adjacent this is to time‑series: rather than forecasting one trajectory, you weight the full space of paths."

It's a genuinely elegant leap. Most forecasting models output a single point estimate — one predicted trajectory, one expected return per asset. The path integral formulation does the opposite: it treats every possible outcome as contributing, with weights that sum to a whole. Applied to markets, that reframes the problem from "what will happen?" to "what is the distribution of things that could happen, and how are they weighted?"

That instinct puts him in good company in this series. Kushpreet made the same move from the energy markets side, abandoning point forecasts for distributional ones because prices have fat tails. Different origins, same conclusion: the distribution carries information that the point estimate throws away.

Vibhav is honest that the idea isn't new.

"Unsurprisingly, it turns out the idea already exists in the literature. In the end my version was a simpler application inspired by this, but leaning on Fourier Analysis rather than the full machinery."

Rather than reaching for the full path integral apparatus, he used Fourier analysis — moving between the time and frequency domain — as his primary lens. It's the tool he says he reaches for most often, describing it as "an interesting perspective to look at price‑series. Nothing crazy, just something I enjoy finding new uses for."

The Ensemble Strategy: Sharpe Plus Novelty

The path integral inspiration became his novelty factor — the part of his submission designed to be structurally different from what other competitors were producing. He then combined it with two other components:

"This was my main novelty factor, which I then mixed with a high‑Sharpe strategy and a time‑decay resistant one — an ensemble that holds Sharpe whilst retaining novelty."

That three‑part construction is worth noting. In AlphaNova's format, a signal has to satisfy two constraints simultaneously: it must be statistically strong enough to enter the quality set, and it must be sufficiently uncorrelated with existing signals to be admitted at all. A high‑Sharpe signal alone can be rejected for redundancy. A novel signal alone can be rejected for weak information content — a fate that befell Alok Verma's deliberately orthogonal submission, and Mathurin's low‑correlation attempt, which was rejected for overfitting because its IC was indistinguishable from noise.

Vibhav's ensemble design addresses both at once. The high‑Sharpe component provides statistical strength. The path‑integral‑inspired component provides structural novelty. The time‑decay‑resistant component provides persistence. It's a portfolio of constraints, not just a portfolio of signals.

For a deeper look at why novelty and diversity matter structurally in a signal library, our research piece on signal diversity, correlation, and the geometry of forecast libraries is worth reading.

Self‑Teaching ML, Properly

When asked about a recent project, Vibhav gave an answer that's refreshingly honest for a competition profile: he's mainly working on his own education.

"Applications have taken up most of my time lately, so my main project is my own upskilling: teaching myself machine learning properly. I've used it plenty in practice but have had no formal training, and my University Physics ML course starts in a month, so this is part pre‑reading and part filling in gaps."

He's working through Financial Edge's Engineering pathway and then mostly YouTube. But the part that stands out is why he's doing it:

"The useful part has been going back over things I'd previously handed to Claude and forming a deeper understanding of why they work."

This is a mature instinct, and it's a counterpoint to the pure acceleration story we've documented elsewhere in this series. Taro demonstrated that you can compete at the top without writing code yourself. Saksham generated 28 candidate signals in two weeks with AI assistance. But Vibhav is pointing at something those profiles don't address: if you outsource the implementation entirely, you can end up with working code and no understanding of why it works. That's fine until something breaks — or until you need to explain your approach in an interview.

AI With Deliberate Separation of Duties

Vibhav uses AI constantly, but he's built a structure around it that's more deliberate than most.

"I handle the idea generation and chip in for problem cases, creative inputs and a final code review. Claude handles everything else, from analysing my approach to implementing the code."

The mechanism he describes is sub‑agents — separate model instances handling different stages.

"For this project I worked in Claude Code, using sub‑agents to keep the stages separate. The separation is what makes it work: a model reviewing something it just wrote tends to defend it, whereas a fresh one offers an unbiased perspective."

That's a genuinely interesting observation, and it maps onto something we've discussed before. Mathurin described the same failure mode from the other direction: language models are exceptionally good at producing plausible narratives, and a coherent story fitted to three data points can survive review precisely because it's internally consistent. A fresh agent with no attachment to the previous output is a partial defence against that.

Vibhav is candid that his current setup is improvised:

"It was fairly ad hoc though, so the next step is organising the workflow properly: a repeatable structure rather than something I assemble from scratch each time."

What We Can All Learn

Vibhav's profile offers three lessons worth carrying forward:

Look outside finance for ideas. Path integrals and Fourier transforms weren't designed for markets. But the questions they ask — about distributions of paths, about behaviour across domains — turn out to be adjacent to the questions signal forecasting asks.
Build ensembles that satisfy multiple constraints at once. A signal has to be strong and novel. Combining components that address different requirements beats optimising for one.
Separate your AI agents by function. A model reviewing its own work tends to defend it. Fresh instances give you an honest read.

And a fourth, perhaps the most important: if you're using AI to build things, go back and understand why they work. Vibhav is doing exactly that, and it puts him ahead of people who never do.

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 Vibhav's did.