About this site

How it was built

The tools on this site are projects I built to understand the models I kept reading about. This page covers how I work, how the public DeanOS differs from the version I use myself, and one example of finding a problem in a model and fixing it. My background is on the home page.

A public rebuild of a personal system

DeanOS on this site is a public rebuild of a system I built for my own portfolio. The personal version also includes a Telegram-based agent for natural-language questions and live broker data; both are left out here for privacy. Rebuilding it in public meant rechecking every model, and some were revised along the way, as the example below shows. The other tools (the Nonlinear Beta Tracker, Options Pricing and the Transaction ML side project) run on the same engine and data.

None of this is a trading system, and I am not a professional quantitative researcher. The tools describe historical data under stated assumptions.

How I work

I use AI coding assistants throughout, the same way many engineers now do, and they wrote much of the code. The work that mattered was mine to do: deciding what each tool should answer, choosing the models and their assumptions, checking results against reference implementations and published formulas, investigating anything that did not hold up, and deciding what to change.

One example: a Value at Risk that was not what it said

DeanOS began as a personal tool. Its Value at Risk offered three methods, one of them labeled Monte Carlo. Rechecking it for the public version showed that the "Monte Carlo" method drew returns from a normal distribution with the portfolio's own mean and volatility, so it was the normal-distribution method again with simulation noise added.

That mattered because it looked like an independent second opinion, when it made the same assumption the normal method makes: that large losses are as rare as a bell curve says. Real daily returns have fatter tails and volatility that clusters.

It was replaced with filtered historical simulation, which standardizes the portfolio's own past returns by their GARCH volatility and rescales them by the next-day volatility forecast, so it keeps their fat tails and responds to current volatility. To check it, the tests confirm it agrees with the normal method when returns really are normal, and every method is backtested on past data: the page counts how often losses exceeded the estimate and applies the Kupiec and Christoffersen tests, which check whether breaches happen at the promised rate and whether they bunch together.

The limits remain: the backtest covers one history, the estimate is for a single day, and the GARCH model has assumptions of its own. The Value at Risk methodology lists them. The same review found two other problems (a volatility forecast labeled as an average that was a single day, and a regime model whose fixed random seed landed on a clearly worse fit); each methodology page lists what changed and why.

How it is built

  • A Python engine (NumPy, pandas, SciPy, statsmodels, arch, hmmlearn, scikit-learn) behind a FastAPI service.
  • A Next.js front end with charts drawn directly in SVG.
  • A nightly job that rebuilds the market data snapshot and precomputes the regime model.
  • An automated test suite, run on every push, that checks the models against exact values, synthetic data with known answers and reference implementations (the tests are public).

What it is not

Nothing here is investment advice or a prediction. The models describe historical data under stated assumptions, and every one of them fails in ways the methodology pages describe. The example portfolios are illustrations, not recommendations. Portfolios you enter stay in your browser's address bar and are never stored.