Ekjot Singh

I study maths at City St George's, University of London, and I'm a quantitative researcher intern at Cabestan Quant Research. I also founded Aleph, a builder community in London.

I'm in the second year of a four-year degree, and I'm applying for summer 2027 internships and 2027 to 2028 placements.

Now, October 2026: entering the Meta Global AI Developer Hackathon with a paper reproducer, and planning a new version of my spurious-factor paper in my own words.

Work

  1. Spurious factor detection under nonstationary covariance

    Jul 2026
    Pilot study

    A standard way to clean a correlation matrix works out the range of values pure noise would produce, treats everything inside that range as random, and keeps anything above it as a real factor. In simulation I found that when volatility comes in bursts and correlations shift once, data with no real factor in it can produce a value about 20 times the top of that range, so the cleaning keeps a factor that isn't there.

    • The largest eigenvalue reaches 44.2, against a noise edge of 2.25.
    • A closed-form formula for the spike is accurate to 1.4% in simulation.
    • On 471 S&P 500 constituents through the August 2011 US downgrade, the formula is within 12%.
    • Portfolios built on cleaning that is fooled this way carry 50 to 65% more out-of-sample risk.
    • Weighting recent data more heavily (EWMA) does not fix it.

    The ideas in this paper are mine; the write-up had AI help.

    A toy version of the paper's experiment. The toy has 48 series over 192 observations with GARCH(1,1) volatility (α = 0.09, β = 0.90), correlation 0 before the midpoint and the slider's value after it, and no factor. The largest eigenvalue is 10.7; the paper's formula predicts 10.1. This is one random draw at toy size; other seeds land between a third below and over half above the prediction, so the numbers above are the paper's, not the toy's. The paper ran the same experiment with 200 series over 800 observations and a shift from 0.05 to 0.45.

    Code and paper

  2. Systematic trading system

    2026
    Active

    A trading system running on an Alpaca paper account. The repository bundles daily SPX, bitcoin and crude oil (WTI) history and the backtest script, so anyone can rerun it.

    It runs behind a kill switch, a circuit breaker, an approval gate and an append-only audit ledger, with a Telegram bot on a local Ollama model.

    The repository's results table, with its out-of-sample split, keeps the crude oil sleeve that lost money.

    Code

  3. PATROL

    2026
    Winner, Captur track

    On-device face blurring for police bodycam footage, built with Devansh Karia at Localhost at Dawn Capital.

    Runs fully offline on the device, with a hash-chained evidence log.

    No public code

  4. Also placed: PropX, winner at the UCL CSRI hackathon; BlockBet, winner at Encode London; SatShield, top 6 at ETH Oxford; LocalFlex, top 14 at the Imperial College London AI Agents Hackathon.

  5. Runway

    Competed

    A cash-flow agent on Xero, built at the Xero App and Agent Hackathon with Encode Club.

    An XGBoost model, Monte Carlo simulation, a seven-guard policy gate and a Xero integration.

    The XGBoost model reached 0.913 AUC.

    Code

  6. Anchor

    Shipped

    A staying-well companion, built at Encode London.

    Next.js and TypeScript.

    It has 32 passing tests.

    LiveCode

  7. AgentSpoons

    A volatility modelling project.

    Code

  8. Aeris

    Mar 2026
    Competed

    A wearable AI patch for paediatric respiratory monitoring, built at MIT Hacking Medicine GrandHack in Boston.

    Code

  9. Paper reproducer

    Oct 2026
    Planned

    A tool that takes a quantitative finance paper, reads the method, writes the backtest and checks whether the claimed results hold up, for the Meta Global AI Developer Hackathon.

    No public code yet

Also competed: ARCBOARD at Encode x Arc DeFi, Overfit Court at the Nebius x NVIDIA Global AI Hackathon, a solo entry at Moth Hack 2026, Syphonix Model to Market, and a UCL real estate hackathon sponsored by PGIM.

Why I do this

I want the things I build to still be useful after the weekend they were built in. When something fails, I leave the failure in the write-up.

Experience

  1. Jul 2026 to now

    Quantitative Researcher Intern, Cabestan Quant Research (Alken Group)

  2. Oct 2025 to now

    Technology Research Analyst, City Innovation Hub, City St George's

    I write articles for the Hub.

  3. 2025 to now

    Undergraduate Representative, London Mathematical Society

    I attend the Society's annual meetings and dinners with academics.

  4. Jul 2025

    One-week placement, Ocado Logistics, Hatfield

  5. Jul 2025

    One-week placement, Publicis Sapient, London

Side quests

Find me

CV available on request

Join Aleph

Aleph has 650+ members as of October 2026. Its events include a Riverside Walk from the Royal Festival Hall to The Anchor at Bankside on 3 October, and Dinner for Ten, an apply-only dinner in mid December.

Join Aleph on WhatsApp

Events are on Luma.