LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
Five projects, not a school name, are what this lecture credits for research and market-making interviews. Four are open source. The trading system is not. Each one forced a modeling choice: which process, which transition matrix, which kernel, which words, which risk limit. The claim is that explaining those choices mattered more than listing a title.
Why it matters
The first project is QFin, a Python pricing library started as an undergraduate and no longer actively extended. It ships arithmetic and geometric Brownian motion and a stochastic-variance process, plus some exotic prices. Install it with pip, import the arithmetic process, pass a volatility, here 30 percent in the demo, and call simulate with a start price, a step count, a time step, and a horizon. The path uses Euler-Maruyama. Interviewers asked about publishing a package and about maintaining one, which most students have not done.
The second project leaves discrete Markov chains for mortgage-backed securities. A Streamlit dashboard loads Fannie Mae data, estimates a transition matrix by maximum likelihood over a lookback, here three months in the walkthrough, and normalizes so each row is a probability. You can ask for the chance of moving from the 30-day bucket into the 90-day bucket over, say, 11 months. An absorbing state and an expected loss are on the page. The lesson the lecture insists on is specification. The chain is a choice. The lookback is a choice. Non-stationarity can make the parameters stale, and the wrong chain is the wrong model. That was enough, in this telling, for serious conversations with portfolio managers.
The idea in plain words
The third project is a Gaussian cookbook, open source, with notes on SSRN. The front page is a Volterra integral, a causal Gaussian process. Recipes cover Brownian motion, a Brownian bridge, fractional Brownian motion, and a Volterra process with a kernel you supply. Fractional paths use a Davies-Harte scheme. The bridge and Brownian motion use a Karhunen-Loève expansion. Each recipe shows the math, the Python, and a continuous or discrete variant. A simulation page lets you add time steps and expansion terms until the empirical covariance approaches the theoretical one. The Hurst parameter smooths a fractional path when you raise it and roughens the path when you lower it. The discussion the project starts, according to the lecture, is not only whether a model can fit a skew. It is whether the fit is fast enough to use. A calibration that takes a day or two is called useless, which is why the lecture keeps pointing at neural-network approximations.
How to use it
The fifth project, the hardest in this account, is a volatility trading system that took a long stretch of design before it placed orders. The public demo does not spray a universe of tickers. It builds one instance for Amazon, sets a position size across levels, the walkthrough uses 5, 1, and 1, and sets a weekly volatility floor. With the share near 244 and implied volatility near 41 or 42 percent, the screen returns an estimated floor of 229, a maximum cash allocation, and a gain per level. Enabling the system sends the order. The project has been rewritten as the author's risk rules changed. The lecture's moral is that an active risk-taker has to understand the model well enough to interpolate and extrapolate, which is why the standing advice is to master the quantitative tools first.
What usually goes wrong
A library nobody can install does not start the maintainer conversation. A Markov dashboard with a single default lookback, and no remark that the matrix might be stale, misses the lesson the mortgage project was built to teach. A Gaussian demo that only shows one pretty path skips the covariance check that justifies the kernel. A blog that restates a textbook without a worked gap does not do the diagnostic job. A trading screen that hides its floor, its size, and its per-level payoff cannot be discussed as risk management.
Questions people ask
What does the pricing library actually simulate?
Arithmetic and geometric Brownian motion and a stochastic-variance process, with Euler-Maruyama paths. The demo uses 30 percent volatility on the arithmetic process.
What data does the mortgage dashboard use?
Fannie Mae performance data. Transitions are maximum-likelihood estimates over a chosen lookback, three months in the walkthrough, including a move from 30 to 90 days delinquent.
What did the Amazon example show?
At a price near 244 and implied volatility near 41 to 42 percent, the system estimated a floor of 229, a cash cap, and a gain per level, then placed the order when it was enabled.