Junior actuary at Marsh Re, with
five years of models in production.
Statistics and econometrics, applied. I work in analytics for global reinsurance broking at Marsh Re in Munich, after five years on bidding and forecasting systems at Holidu (~€100k/day in ad spend, 24 markets) and a year of rate studies and financial forecasts for municipal water and sewer utilities. Two probabilistic forecasting models of my own ran in public, scored against market consensus. Munich · DAV track · bilingual English and German.
Four projects. Method, outcome, and role for each.
Two at Holidu, one for municipal water and sewer utilities, one academic.
Two forecasting systems, run in public and scored against the market.
Built and run end to end — data pipeline, model, scheduled agent, and dashboard. Every prediction was logged before the event and scored afterwards against market consensus. Both dashboards are still up, with the full record.
World Cup 2026 prediction model.
Forecast all 100 matches of the tournament, each one logged before kickoff and scored afterwards: log-loss 1.003, ahead of a national-Elo baseline (1.010) and well ahead of a coin flip (1.099) — but the vig-free market finished ahead at 0.854–0.889 and won 64 of the 100 head-to-head matchups, with the knockout rounds level. Player-level strength ratings rolled up to team ratings, a 10,000-run Monte Carlo simulated every match and the full bracket, and a matchday agent conditioned each forecast on the confirmed lineups at kickoff. Pipeline, model, agent, and dashboard all self-operated; the site is now a retrospective of the full record.
March Madness bracket model.
Scored out of sample on 11,170 games across two complete seasons: 71.6% accuracy, Brier 0.183, AUC 0.772. On the 4,240 games that also carried market odds the market stayed narrowly ahead (71.8% and Brier 0.183, against 70.2% and 0.192) — competitive, not market-beating, and the dashboard says so. Underneath: an XGBoost win-probability model over 16,691 NCAA Division I games with leakage-free feature engineering, walk-forward evaluation, 10-bin calibration and SHAP attribution, plus a three-layer player-availability system and a 100k-run Monte Carlo bracket simulator.
How I work.
My background is statistics and econometrics; I ship the model as software the team runs.
Method follows the question, not the résumé: GLMs for most of it, distribution fitting and bootstrap when uncertainty is the answer, time-series and tree-based models when the data asks for them.
I've worked where the gap between a good forecast and real money spent is short — performance marketing, municipal finance, reinsurance, and forecasting systems I've built and run, some solo and some with a team. The models get built to explain themselves to the person whose decision they inform, and the manual work around them gets automated: at Holidu that cut the weekly market-by-market routine roughly in half.
I'm bilingual in English and German, based in Munich, and work in analytics for global reinsurance broking at Marsh Re. Alongside that I'm studying toward actuarial certification on the DAV track — pricing, reserving, and the statistics underneath them.
Experience & education.
The short version — experience, education, and what I'm working toward.