Curriculum · Pillar Four

AI & quantitative
finance

Where machine learning genuinely meets modern markets — the same material students worked through in a Python notebook at our Finance & AI Workshop.

Quantitative finance is the use of mathematics, statistics, and code to make financial decisions. Machine learning is now part of that toolkit — but not in the way most coverage implies. This page describes what these methods actually do, and where they fail.

What a quant actually does

The work is mostly data, not prediction. A typical project is: obtain a dataset, clean it, construct features that might carry signal, fit a model, and then test it honestly on data the model has never seen. The final step is where most ideas die, and where the discipline lives.

Roles cluster into researchers who look for signal, developers who build the systems that trade it, and risk teams who size positions so a bad period is survivable. All three write code; the working language is overwhelmingly Python.

Where machine learning helps

  • Pattern recognition across many inputs. Models can weigh hundreds of variables at once in ways a human cannot hold in their head.
  • Unstructured data. Language models can turn filings, transcripts, and news into structured numbers — sentiment, topic, change relative to last quarter.
  • Operational work. Fraud detection, credit scoring, and trade execution are the areas where these methods are unambiguously in production and unambiguously effective.

Where it fails — the part that gets skipped

Financial data is unusually hostile to machine learning, for reasons worth naming precisely:

  • The signal-to-noise ratio is terrible. Price movements are mostly noise. Flexible models fit noise beautifully.
  • Overfitting looks like success. Test enough strategies on the same history and some will look brilliant by chance alone. A backtest is a hypothesis, not a result.
  • The system adapts. A genuine edge, once traded at scale, tends to disappear — because other participants find it too.
  • Lookahead bias. Accidentally using information that was not available at the time is the most common and most invisible bug in the field.

None of this is a reason to avoid the subject. It is the reason serious practitioners spend more effort trying to break their own results than to produce them.

The notebook we ran with students

At the Finance & AI Workshop, students worked through a Python notebook that built a simple stock model end to end. The point was not the model's accuracy — it was seeing every step made explicit:

  1. Load historical price data with pandas.
  2. Compute returns and a handful of simple features, such as moving averages.
  3. Split the data by time — training on the earlier period only.
  4. Fit a basic regression with scikit-learn.
  5. Evaluate on the later, unseen period and compare against simply holding the asset.

The lesson most students took away is the honest one: a first model usually does not beat buying and holding, and understanding why teaches far more than a model that appears to work.

How to start, concretely

  1. Learn Python properly — functions, loops, and data structures before anything financial.
  2. Learn pandas. Most quantitative work is data manipulation.
  3. Take statistics seriously: distributions, regression, and what a p-value does not mean.
  4. Build one small project end to end and try hard to disprove your own result.
  5. Keep linear algebra and calculus alive; they underpin everything later.

Next

This is the last of our four pillars — start from the curriculum overview, or read about the Finance & AI Workshop where we taught it.