QuantPad

What it is

A CodinGame-style platform where you train for a quant, data or M&A seat by writing code that is graded as you type. Lessons end in a puzzle you must make pass, interview cards come back in spaced repetition, and every role shows how much of its route you have validated.

The models behind the puzzles

169 puzzles and 804 test cases, difficulty 1 to 6, written so that the maths has to be implemented, not quoted:

• Pricing and volatility: Black-Scholes and Greeks, binomial trees, Monte Carlo with variance reduction, barriers, SVI smiles, GARCH term structure.
• Rates and credit: curve bootstrapping, Vasicek, yield-curve PCA, CDS survival curves.
• Statistical arbitrage: cointegration, Kalman filters and Ornstein-Uhlenbeck spreads, $d{X}_t=\theta (\mu -X_t)dt+\sigma d{W}_t$.
• Execution and market making: Almgren-Chriss, which minimises $\mathbb{E}[C]+\lambda \mathrm{Var}[C]$ and trades along $x_j=X\sinh (\kappa (T-t_j))/\sinh (\kappa T)$, and Avellaneda-Stoikov, with reservation price $r=s-q\gamma {\sigma }^2(T-t)$.
• ML for trading: rank IC, purged cross-validation, bet sizing and the deflated Sharpe ratio, so a backtest is judged against the number of trials that produced it.

The SVI smile, as a learner meets it

The Visual Lab has 32 interactive figures of this kind: delta hedging one path day by day or 300 paths at once, Monte Carlo convergence, VaR and ES under fat tails, the efficient frontier, Kelly growth.

Grading floats honestly

Quant answers are floats, so an exact check would fail a correct solution that used another erf approximation. Answers are compared leafwise with a combined tolerance $|x-y|\le {\epsilon }_{abs}+{\epsilon }_{rel}|y|$. Test vectors are produced by running each puzzle's own reference solution, then replayed as a self-check: a puzzle cannot ship with a test its solution fails.

Engineering

• User code runs in a Web Worker with a quant standard library injected; a hard timeout terminates the worker, the only reliable way to stop an infinite loop.
• JavaScript, TypeScript and Python (Pyodide) run in the browser; Julia, C++, C# and OCaml on a hosted runner, through a harness generated from each puzzle's test vectors.
• Duel bots draw solve times from a log-normal fitted to difficulty and skill, ratings follow Elo.
• Backtests only ever hand the strategy past closes, so look-ahead is impossible by construction, with an in-sample / out-of-sample split.
• Proctored exams record eleven integrity signals (keystroke cadence, paste share, structural similarity) into a SHA-256 hash chain.

Role

Designed and built end to end: content, grading engine, runtime, design system, bilingual EN / FR.