I’m Christopher Ludwig, an industrial engineering student at RWTH Aachen. I build models and measuring tools, then try hard to break them.
Who am I?
A student of industrial engineering and management (mechanical engineering track) at RWTH Aachen, and a research assistant at the Surface Engineering Institute, where I measure plasmas in coating processes. Before university I worked in production at PAVOY and in business development at the EdTech startup SIMPLEX EDUCATION.
Outside of that I play jungle for RWTH’s League of Legends team in the German collegiate league, play chess, lift, and learn Japanese.
Projects
Drag the cube to turn it. Click a side to open it.
All sides are listed below.
Backtest Lab
A research engine for trading strategies, written in Rust with a browser interface. It searches thousands of rules, tests them walk-forward, keeps the final years locked away until the end, and paper-trades the survivors through a broker API.
After realistic costs, nothing beat buy-and-hold by more than random rules do.
This curve is lying
It is the best of 1,000 trading rules, picked on three years of random prices. Nothing in that data can be predicted. Run it on two years it has never seen.
…Sharpe where it was picked
?Sharpe on unseen data
Sharpe ratio: return per unit of risk, per year. On random prices every rule is truly worth zero. Usually the winner falls back toward it. Sometimes luck holds for a while, and that is the trap.
25Markets in the trend-following test
3Accounting bugs found by an audit, all fixed
22Regression tests added for those fixes
−0.16 to −0.20Walk-forward Sharpe versus buy-and-hold
What it does
Timing strategies on single assets, stock picking across a portfolio, pairs trading, and intraday setups such as opening-range breakouts and VWAP rules.
Mass search over thousands of parameter sets, next to random baselines that show what luck alone produces.
Walk-forward optimisation and a locked holdout that is opened once, at the end.
Paper trading through the Alpaca broker API.
What went wrong, and how it was fixed
An independent audit found three accounting bugs that made results look better than they were: phantom rebalancing, exposure caps that did not hold, and fills that used prices from the future. I rebuilt the execution core as one shared kernel; 22 regression tests pin the fixes down, and the results were cross-checked against a second, independent implementation.
What I tested
Price gaps between share classes of the same company, holding index funds only overnight, the drift before Federal Reserve meetings, trend following across 25 markets, and exits based on risk multiples. None kept a profitable edge out of sample once realistic costs were in.
Built with heavy AI assistance. The research questions, test rules and decisions are mine.
Racing model
Win probabilities for German horse races from form, jockeys, trainers and the betting market itself, tested out of sample, with costs, and against luck.
A hair better than the market at predicting winners. Still clearly unprofitable to bet.
How much can 125 bets tell you?
A bettor with no edge at all: every bet fairly priced, minus the tote’s cut. Each bar counts simulated seasons by their return; pink bars made money anyway.
…Seasons in profit despite no edge
…Middle 90 % of seasons
3,000 simulated seasons per setting, odds between 2.5 and 25, cut set to the median overround of 1.19 in the race data.
2,465Official races, Jan 2024 to Oct 2026
20,676Starters
15.4 %Holdout pseudo-R², my blend (market 15.2 %)
−20.6 %Holdout return of the frozen strategy, 125 bets
Models, holdout of 372 races
Model
Log loss
Pseudo-R²
Blend of my model and market odds
1.736
15.4 %
Market (final tote odds)
1.740
15.2 %
Conditional logit, form only
1.908
7.0 %
LightGBM, form only
1.919
6.4 %
Elo rating only
1.984
3.3 %
Every horse equally likely
2.051
0
Betting, holdout, flat 2 € stakes
Strategy
Bets
Return
Frozen strategy: blend, edge of at least 5 %
125
−20.6 %
Always the favourite
369
−13.3 %
A random runner
369
−23.3 %
The strategy was fixed and committed before the holdout was opened. Every model is refitted monthly on earlier races only; the blend follows Benter (1994). 49 automated tests cover the pipeline.
Built with heavy AI assistance. The research questions, test rules and decisions are mine.
Plasma probe
Hard coatings on tools and engine parts are made in plasmas. How many free electrons that plasma holds shapes the coating, so it is worth measuring while the process runs. That is part of my job at RWTH’s Surface Engineering Institute (IOT).
Ongoing research. This explains the method; results stay with the institute until they are published.
Read the density from the resonance
The probe sends a microwave sweep into the plasma and listens for the frequency it absorbs. That resonance moves with the electron density.
…Plasma frequency
…Probe resonance
Idealised probe without a sheath: resonance = plasma frequency ÷ √3, plasma frequency ≈ 8.98 kHz × √(electrons per cm³). An illustration, not measured data.
Why count electrons
In physical vapour deposition (PVD), ions from the plasma strike the layer as it grows. The plasma density sets how many arrive, which changes how dense and hard the coating becomes. HiPIMS pushes this further with short, very intense pulses.
Why this probe
The multipole resonance probe sits in the chamber as a small sphere and needs only a microwave signal. Its electrodes are covered by a dielectric, so insulating films that ruin many other probes do not stop it.
Experience
Research assistant, Surface Engineering Institute (IOT), RWTH Aachen
B.Sc. Industrial Engineering and Management, RWTH Aachen
Production intern, PAVOY GmbH, Hillesheim
Business development intern, SIMPLEX EDUCATION GmbH, Aachen
Production intern, PAVOY GmbH, Hillesheim
Skills
Languages: German (native), English (fluent), Japanese (beginner, A1 course at RWTH), Latin (basics).
Tools: Python (basics), AI-assisted development in Rust and Python, MS Office, Adobe Photoshop.
Methods: out-of-sample testing, walk-forward evaluation, market and competitor analysis.
Get in touch
Internships, research positions, project work. Write here, or email kontakt@studioludwig.de.