The Minds Behind the Models: Saksham Arora
Saksham Arora: Deliberate Differentiation and the Discipline of Tail Latency
Saksham Arora doesn’t think like the average participant. His background isn’t purely finance or machine learning—it’s a blend of low‑latency C++ engineering, data pipeline construction, and a decade of training as a Grade 8 drummer (Trinity College London). He’s currently navigating a dual‑track academic path at IIT Madras and Guru Gobind Singh Indraprastha University, while shipping projects that span from competitive programming benchmarking to an LLM latency measurement tool. In AlphaNova’s Competition 5, he applied a software engineer’s obsession with measurement and a musician’s sense of rhythm to the problem of generating trading signals—and the result was a deliberate, counterintuitive choice that moved him from rank 167 to rank 6.
His path into quant finance began, like many, in a competition—IMC Prosperity 4.
“I went in with two friends, mostly curious, and we ended up ranked 154 out of 18,800+ teams. What hooked me wasn’t the ranking, it was realising that markets are a place where you can be measurably right or wrong.”
That clarity—a system where correctness is binary and verifiable—resonated with his engineering mindset. And it carried directly into his AlphaNova strategy.
One of the central challenges of AlphaNova’s competition is that signals are scored not just on their out‑of‑sample Sharpe ratio, but on their novelty relative to the existing pool. A brilliant signal that’s highly correlated with already‑accepted submissions is skipped. So Saksham made a deliberate trade‑off: he sacrificed raw performance to reduce correlation.
“The main decision I made was to deliberately make my signal less correlated with what everyone else was submitting. That cost me about ten percent of my Sharpe, from 0.0537 to 0.0481, and moved me from rank 167 to rank 6.”
It’s the kind of move that feels wrong on a standard leaderboard—punishing yourself in a metric you’ve been trained to maximise—but it was exactly right for this environment. Saksham understood that the scoring function was not just accuracy, but uniqueness. And he acted on it.
That instinct for measuring the right thing shows up in his side projects, too. When asked what he’s been excited about lately, he pointed to llm-bench, a small CLI tool he built to benchmark LLM inference latency properly.
“Most benchmarks quote averages, which hides everything that matters. I built it to report TTFT and inter‑token latency at p50, p95, and p99 across providers, because tail latency is the number you actually care about if you’re building on top of these APIs.”
The project surfaced a bimodal latency distribution in a live benchmark—where the mean time‑to‑first‑token was lower than the median, revealing two completely different performance regimes hiding inside a single number. It’s the same kind of statistical skepticism that makes a quant question a backtest with an overly smooth equity curve.
In his AlphaNova workflow, Saksham applied a similar principle: don’t trust the aggregate. Trust the recent.
“Testing a signal on the most recent slice of data rather than the whole history saved me from submitting things that only looked good in a backtest. My full‑sample results looked much better than what actually showed up out‑of‑sample, and the last third of the data was a far better predictor of the real score.”
It’s a simple check, but one that’s easy to skip when you’re chasing the highest possible backtest metric. For Saksham, it became a filter that eliminated false positives before they ever reached the platform.
AI tools played a significant role in his process—not as a replacement for thinking, but as a force multiplier for experimentation.
“In this competition I went through something like 28 candidate signals across a couple of weeks, which I could not have done by hand alongside everything else. So AI widened the search rather than replacing the thinking. The rule I hold to is that nothing enters my results until I have re‑run it myself.”
That distinction—between generating ideas and validating them—is one we’ve heard from several top participants. AI accelerates the exploration, but the judgment, the testing, and the final call remain human.
Saksham’s broader interests reinforce his quantitative edge. His experience as a drummer taught him to think in layered, polyrhythmic structures, and his work at Airtel and Cimplifie sharpened his ability to build scalable data pipelines and ETL systems. He won the Paytm Hackathon 2026, competed in the NASA Space Apps Challenge and the Agentic AI Hackathon, and shipped the LLM benchmarking tool to PyPI. For him, quant finance is another domain where rigorous measurement and systematic thinking yield an edge.
You can find Saksham’s work at saksham.digital and explore the llm-bench tool on GitHub.
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 Saksham’s did.