The Minds Behind the Models: Alok Verma

Alok Verma: The Mechanical Engineer Who Built His Own Evaluator

Alok Verma started out studying mechanical engineering at NIT Tiruchirappalli. Today, he's a Lead Data Scientist at Kentrix AI, a Master‑level WorldQuant consultant, and one of the sharpest infrastructure thinkers in our Competition 5 cohort. His path into quantitative finance wasn't planned — it emerged from curiosity, competition, and a habit of building systems that tell him when he's wrong.

"It all started during my undergraduate years at NIT Tiruchirappalli. As a mechanical engineering student, I realized fairly early that I wanted to explore beyond my core discipline. I began experimenting with software development on my own, and as AI and machine learning were evolving rapidly, I became increasingly interested in data, how we could use it to identify patterns, build models, and make better decisions."

Midway through his second year, he came across WorldQuant's International Quant Championship. He had virtually no background in finance or quantitative trading — but the competition appealed to him because it was fundamentally a problem‑solving challenge. He learned the concepts from scratch while competing.

"The 2024 championship became a major milestone. I achieved Global Rank 4 in Stage 1 and Rank 1 in the opening round of Stage 2, eventually reaching the National Finals as one of eight teams from more than 34,000 participants. Coming from a mechanical engineering background, that experience gave me the confidence that I could build expertise in quantitative finance through independent learning and rigorous experimentation."

That journey led him to the WorldQuant BRAIN platform, where he has developed more than 1,500 long‑short market‑neutral alphas since May 2023, held Master level for three consecutive quarters, and ranked 32nd globally and 8th in India in the MAPC Star Competition among over 1,000 research consultants. He also competed in IMC Prosperity, finishing in the top 5% among 14,000+ participants.

"Looking back, the path was completely unplanned. I moved from mechanical engineering to software, then machine learning, and eventually quantitative finance."

The 0.52x Calibration That Changed Everything

Alok's standout contribution to Competition 5 was a piece of infrastructure: a fast local replica of the official evaluator.

"For AlphaNova, I built a walk forward harness that reproduced the official scoring pipeline, allowing me to evaluate a candidate signal across all 77 validation periods in seconds rather than using a submission to find out whether it worked."

But the real insight came from calibrating that replica against reality.

"One of my flagship signals scored 0.077 locally but 0.0403 on the server. Instead of treating the difference as noise, I treated it as data. I found that my local Sharpe mapped to the server at roughly 0.52x and that the hidden test window behaved more like my later validation periods. That led me to use last 30% Sharpe as my selection metric, and my subsequent submissions landed very close to their predicted scores."

This is the same insight Mathurin Aché Arnaud, a data scientists from France, arrived at independently — and it's becoming a recurring theme in these profiles. When your validation number diverges from the server number by a consistent factor, the divergence itself becomes a signal. Optimising the raw validation Sharpe is optimising the wrong thing.

"So the tool became more than a backtest. It became a feedback loop. Every submission made the evaluator more accurate. The models themselves were often simple Ridge regressions. The real edge came from the infrastructure around them."

Simple Models, Serious Validation

Alok deliberately kept his models simple. His main approach was Ridge regression on cross‑sectional ranks, focusing on the relative ordering of assets rather than their raw values. Everything was evaluated walk‑forward, with training restricted strictly to information available at the time.

"The real work was around the model: feature selection, validation, calibration against the hidden evaluator, and understanding what actually generalized. I used all ten submission slots and finished with a best test Sharpe of 0.0434 and a quality set star."

His most valuable lesson came from a failed submission.

"I deliberately built a signal that was different from the crowd and genuinely causal, but I had sacrificed predictive strength in order to make it more novel. The server rejected it as overfit because its information coefficient was too weak to distinguish the signal from noise on unseen data."

That experience crystallised a rule he now follows closely:

"I would rather have a statistically significant edge that is similar to something others have found than a novel idea that cannot survive out of sample."

He also noted a counterintuitive finding that reinforces something we've seen across the competition:

"My most predictive submission by IC actually produced negative Sharpe. A modest but stable edge, combined with disciplined validation and low turnover, proved far more useful."

"Consistency matters more than complexity."

Building Systems That Can Tell You You're Wrong

Outside of AlphaNova, Alok is building FinOps Brain, a product in the Indian tax and compliance space. It's a reconciliation problem — matching large, messy financial datasets that describe the same transactions but differ in format, naming, and completeness.

"Since every mismatch can have a monetary consequence, I designed it around a simple principle: be deterministic wherever possible and never silently guess."

The system uses strict validation, idempotent processing, versioned transformations, exact decimal arithmetic, and an immutable audit trail. When it can't confidently resolve a mismatch, it escalates to a human with the relevant evidence rather than forcing an answer. The LLM is kept out of the deterministic path entirely and used only to assist with review.

"Building this has taught me that good engineering is often less about making a system clever and more about making it predictable, auditable, and difficult to fool. That mindset has also influenced how I approach quantitative research. The validation layer deserves as much attention as the model itself."

He's also built portfolio optimisation and forecasting pipelines using ARIMA and ETS models, combined with risk analytics including VaR, CVaR, and PCA‑based measures of portfolio concentration.

"The common theme across both projects is the same: build systems that can tell you when you're wrong."

AI as a Force Multiplier, Not a Replacement for Judgment

Alok uses AI extensively — but with a clear boundary.

"During AlphaNova, I used Claude Code as an engineering and experimentation assistant. I would define a hypothesis, and Claude would help implement it within my walk forward framework, allowing me to test ideas across all 77 validation periods quickly. In one evening, I tested nine variants of a flagship signal. Only one improved the results, and I pursued that further."

He also uses AI for repetitive tasks like coding, running experiments, and documenting results, which frees him to spend more time on research itself.

"What I don't delegate is hypothesis selection or result interpretation. Language models can produce very convincing explanations for noisy results, so I rely on rigorous validation and holdout testing to decide what is actually real."

"For me, AI has made quantitative research faster and more scalable, but the judgment that decides what to test, what to trust, and when to say no remains human."

What We Can All Learn

Alok's profile highlights three lessons that apply well beyond Competition 5:

Calibrate your local evaluator against the real one. If your validation Sharpe maps to the server at a consistent factor, use that factor. Don't optimise the wrong number.
Simplicity plus discipline beats complexity. Ridge regression on cross‑sectional ranks, evaluated rigorously, produced a quality set star. The infrastructure around the model mattered more than the model itself.
A novel idea that can't survive out‑of‑sample is worthless. Alok learned this the hard way — and now prioritises statistically significant edges over decorative originality.

That combination of engineering discipline, honest calibration, and intellectual humility 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.