Second-year BSc Economics, Finance and Data Science, Imperial College London, on track for First Class Honours
Alex Lahdiri
I build systematic trading strategies and machine-learning systems, and I write research on the Riemann Hypothesis. I treat every promising result as guilty until proven robust.
- Next
- Quantitative Trading Intern, Optiver, Amsterdam. Incoming, summer 2027.
- Research
- Two sole-authored working papers in analytic number theory. Not yet peer reviewed.
- Recognition
- Dean's List 2025/26, Imperial Business School. Top 10% of the cohort.
Fig. 1. Nine Gaussian wave packets, the test functions in my research, each drawn as the curve u ↦ (u, Re h, Im h): a helix under a Gaussian envelope. Move the pointer to turn them.
About
Over the past year I have written two working papers on the Riemann Hypothesis, built an AI trading desk that finished 31st of 400+ in Model to Market on a simulated book, written my team's algorithms for IMC Prosperity 4, and worked on Responsible AI governance with Fundación Paraguaya.
I'm now in my second year at Imperial, continuing the research alongside my degree, and I join Optiver as a quantitative trading intern in summer 2027.
2026, sole author
Weil positivity and the Riemann Hypothesis
How much of the Riemann Hypothesis can be checked with finite matrices?
Mathematics, not simulation. Proofs supported by numerical experiments. Working papers, not peer reviewed. Neither paper claims a proof of the Riemann Hypothesis.
The question
The Riemann Hypothesis is equivalent to a certain quadratic form being positive. I wanted to test that positivity locally, with Gaussian wave packets, and find out exactly where it becomes hard.
What I proved
Paper I shows the whole question reduces to checking that finite matrices built from those packets are positive. It proves the reduction unconditionally, together with a scale law and almost-all positivity, supported by nine numerical experiments, and it isolates dense clusters of packets as the remaining obstruction.
Paper III takes on that case. It resolves dense clusters up to a natural ceiling set by the number of zeta zeros nearby, and shows that going beyond that ceiling is equivalent to the Riemann Hypothesis itself.
Paper III uses a coherent-state (Fock-basis) expansion and a matrix moment inequality built on the non-commutative Khintchine inequality.
- closest pair
- –
- Gershgorin bound
- –
- λmin
- –
- J
- –
- zeros in window N
- –
- ceiling
- –
The decision that mattered
Withdrawing my own results.
Version 4 of Paper I withdraws two lemmas and the proofs of two theorems, reclassifies part of the earlier argument as heuristic, and drops a claim I had made under the Lindelöf Hypothesis. The version note at the front lists the changes. A smaller statement I can defend is worth more than a larger one I can't.
The outcome
Two working papers, under the name A. Ruzzene-Lahdiri.
As a by-product, Paper III proves an unconditional lower bound on the number of zeta zeros in short windows at almost all heights. There is a Zenodo record, and the manuscripts are available on request.
- Paper I (working paper, v4). Pole-free Gaussian wave packets and a finite-dimensional reduction of the Weil positivity criterion. Paper II was merged into it in version 4.
- Paper III (working manuscript). Dense clusters, the zero-count ceiling, and a matrix moment inequality.
June 2026, Syphonix × AI Engine competition
Model to Market
An AI desk proposes the trades. A rulebook it can't overrule decides the size.
Simulated. A $1,000,000 paper account trading live prices through MetaTrader 5. No real money was at risk.
The question
Can a language model help run a trading book without being trusted with its risk? Every entrant had the same simulated account and live market data, scored 70% on return, 15% on drawdown, 10% on Sharpe and 5% on risk discipline.
What I built
The whole stack. A pipeline for a 20 GB tick archive, a backtester with per-instrument costs, and a market-neutral book across eight FX pairs plus gold and silver, sized by Ledoit–Wolf shrinkage mean-variance and executed through MetaTrader 5.
Above it sits an AI desk. NVIDIA Nemotron classifies the live regime, Claude proposes the strategy and size as a validated object, and a deterministic governor has the final say. You can be the AI below.
You are the AI. Ask for a bigger book.
Size is gross exposure: the total value of all positions, as a multiple of the account.
Or run the three cases in my test suite
3.0×
$3.0M of positions on the $1,000,000 simulated account. No single position above $750k.
The safe band binds. Whatever the model asks for, gross exposure stays between 1.0× and 3.0×. A language model can be confidently wrong, so the limit never depends on it behaving well.
The rules, in the order the code applies them
- Edge testSignificant edge (live IC above 0, t-statistic at least 2): no cap.9.0×
- Drawdown checkLess than 5% below the peak: no cap.9.0×
- Safe bandAbove the band: cut to the 3.0× ceiling.3.0×
- Three-minute monitorDrawdown ladder: up to 3% down, full size.3.0×
The design rule. The AI can propose anything. Only deterministic code sizes the book, the monitor can only cut it, and the limits are covered by a test suite I can run.
The decision that mattered
Retiring my own best strategy.
My backtest champion was momentum. On the live feed, an information-coefficient and autocorrelation study showed it had the wrong sign: hourly returns were mean-reverting, with lag-1 autocorrelation around −0.1 to −0.2. I retired momentum mid-competition and built a mean-reversion strategy in its place.
The overfitting checks had already told me not to trust the size of the backtest.
The outcome
31st of 400+, and a live finalist.
In the full-month backtest the book had zero forced liquidations and never entered a margin, leverage or concentration penalty tier. The edge was thin and regime-dependent, and I said so in the submission. I also published what didn't work: momentum, a reversal sleeve that died to costs, crypto, order-book signals and scalping.
April 2026, team of four
IMC Prosperity 4
Market making, options and Monte Carlo, against 18,803 teams.
Simulated. IMC's trading competition: a simulated exchange and in-game currency.
The question
Each round added new products to a simulated exchange, from plain order books to options and exotics. The question was the same every time: where is the edge, and is it bigger than my own measurement error?
What I built
My team's algorithms, end to end. A market maker on a mean-reverting product (Ornstein–Uhlenbeck, half-life about 3 ticks) and a trend follower on one whose drift fitted with R² = 0.99999. Black–Scholes implied-volatility smiles across a ten-strike option chain, where the round's PnL rose 101% over 15 iterations. Ten exotic options priced by Monte Carlo with 500,000 paths and antithetic variates, and a sealed-bid auction solved by expected-value search.
Along the way I tested Kalman filters, Avellaneda–Stoikov quoting, Hawkes processes and hidden Markov models.
The decision that mattered
Trading five of ten.
I priced ten exotic options but traded only the five whose edge was larger than the simulation's own noise. A price difference smaller than your pricing error isn't an edge.
Traded: five edges above the noiseLeft alone: five inside it
The outcome
1st worldwide in Round 1 manual trading. Top 15% of 18,803 teams overall.
May to June 2026, first-year project at Imperial
Learning to collude?
Can you tell from prices alone that pricing algorithms have stopped competing?
Simulation, then real data. Simulated pricing agents, tested on live London diesel prices.
The question
Pricing algorithms that never communicate can still learn to keep prices high. I wanted to know whether a classifier trained on simulated markets could recognise that pattern in real ones.
What I built
The machine-learning pipeline; a teammate helped with the economics. Two Q-learning agents play a repeated Bertrand pricing game with logit demand. A markup index scores each run, a gradient-boosting classifier learns to separate collusive from competitive price paths, and a live pipeline polls London diesel prices every six hours to test it.
Each run is scored by its markup index
The decision that mattered
Not believing 0.95.
The first classifier scored a ROC AUC of 0.95. I went looking for the reason and found data leakage. I fixed it and retired the number. I didn't record the score after the fix, so I don't quote one.
The outcome
A working simulator, classifier and live data pipeline, presented as a poster.
What I'd stand behind is the method, not a headline accuracy.
Shorter pieces
-
QRT C3 Hackathon, October 2026. Scored 85% on the black-box optimisation question.
Optimised an unknown function from noisy evaluations by probing it with targeted queries and reverse-engineering its structure.
-
Rolls-Royce long thesis, Citadel Global Equities Competition, January 2026. Target £14.95 against a £12.04 share price (+24%).
A valuation, not a trade. Argued the market prices Rolls-Royce as a cyclical engine maker when 66% of its civil aerospace revenue is recurring services, with base and thesis DCFs, a sum-of-the-parts check and sensitivity grids. Team of four.
CV
The short version. The detail is on LinkedIn.
Experience
- 2027Quantitative Trading Intern, Optiver, Amsterdam (incoming, summer 2027)
- 2026AI Researcher (Intern), Fundación Paraguaya, Responsible AI Centre, May to JulyRisk, ethics and governance framework for AI used with the Poverty Stoplight. Team of four.
- 2026Dive into Quant Strats, Goldman Sachs, London, AprilInvite-only insight day with AmplifyME. Captained a pair that ran in the top 5 of about 40 in a live ETF arbitrage simulation.
Recognition
- 2026Dean's List, Imperial Business School, top 10% of cohort
- 2026IMC Prosperity 4, 1st worldwide in Round 1 manual trading
- 2026Model to Market, live finalist
- 2026Y Combinator Startup School, Paris, selected participant
Education
- 2025–28Imperial College London, BSc Economics, Finance and Data Science. On track for First Class HonoursProbability and Statistics, Econometrics, Machine Learning, Data Structures and Algorithms, Corporate Finance and Capital Markets. Algorithmic Trading, Mathematics, Computing, AI and Investment societies.
- 2025Clifton College, A-Levels: A*A*A*A* (Maths, Further Maths, Physics, Economics), A* EPQDeputy Head of School. Captain of debating: school champions 2025, Hugh Harper Prize for best debater.