Quant researcher at NTU Singapore, builder of validated data and forecasting systems. I tested 13 market anomalies — 12 died. One survived: S&P 500 index additions, +2.27% per event, t-stat 2.93 across 102 events. I build the same way: validation before belief, uncertainty made explicit, everything auditable.
Four principles run through every project. They are also why my models are smaller than they could be — I reject what I can't defend.
Every idea is assumed wrong until real data says otherwise. Most don't survive — and that's the point.
Leak-free, out-of-sample, walk-forward. A model earns its place by beating a baseline, or it's cut.
Point estimates lie. I ship intervals and label what they don't cover. No false precision.
Every number traces to a source. Reproducible offline. Tests gate the claims.
A live probabilistic World Cup forecast: Elo→Dixon-Coles + an ML ensemble, multi-provider live xG and odds, validated leak-free on four past World Cups, with champion-probability intervals.
Systematic signal research on SEC EDGAR filings. Structural, forced-flow edges over behavioral noise — the S&P 500 index-addition effect ($15.6T in tracker funds) is the anchor case.
An all-in-one platform for French law students: AI document classification, adaptive QCM, an AI editor with correction, and unified legal search across Legifrance / EUR-Lex / CEDH.
A private investment-club operations system: organized data base, automated reporting pipelines, and strict confidentiality controls. Process design over hype.
A research harness for testing market anomalies end-to-end: data, hypothesis, significance testing, and an honest reject/keep gate. The infrastructure behind the kill-tests.
Most World Cup "predictions" are confident guesses. This is a probability distribution built to survive a hostile review — not to look impressive.
Live data from five providers, cross-checked by a disagreement system:
Honest caveats, stated on the project itself: the xG providers likely share an upstream (not independent); the intervals are a floor (parameter sampling only, not total uncertainty); a dynamic upset-robust ML mode exists but is not the default. It's a forecasting lab, not a calibrated guarantee.
I'm Yorian Melki, a quant researcher and systems builder. I care about the part most people skip: validation, uncertainty, and reproducibility. A model I can't defend under attack doesn't ship.
Right now I'm at NTU Business School Singapore (Simulation Techniques in Finance), building public proof-of-work across forecasting, equity-signal research, and product. The thread is the same everywhere: structure over story, evidence over confidence.
I'm heading to HSG St. Gallen next, continuing on a quant / research trajectory — and compounding a portfolio of real, auditable projects along the way.
Open to quant research, data/AI systems, and serious collaboration.