Researcher course in machine learning
Financial Machine Learning: Deep and Reinforcement Learning
The research methods developed for machine learning on financial data: sampling with information-driven bars, triple-barrier labels and meta-labeling, sample weights for overlapping outcomes, purged cross-validation, feature importance, and statistics of backtest overfitting. It also teaches deep learning for finance and reinforcement learning for trading.
Machine learning methods were built for data with many independent samples and a stable relationship to learn. Financial data has neither. Over the past decade, researchers have developed financial machine learning: methods specific to finance, for how to sample, label, weight and validate, and how to judge whether a backtest is the product of skill or of many trials. This course teaches those methods at research level.
What is the triple-barrier method?
A way of labeling trading outcomes that matches how a position is actually managed. From each entry point, three barriers are set: a profit target, a stop-loss and a time limit. The label records which barrier is touched first. It produces labels that reflect real trade outcomes rather than an arbitrary fixed-horizon return.
How do you tell if a backtest is overfit?
By accounting for how many strategies or settings were tried. The deflated Sharpe ratio adjusts a Sharpe ratio for the number of trials and the shape of the return distribution, and the probability of backtest overfitting estimates how often the best in-sample choice underperforms out of sample. You will apply both to your own research.
Why do so many deep learning trading results fail?
Deep learning for finance applies neural networks to prices, order books, fundamentals and text. It has produced real gains in some places and a great many false ones. This course teaches the models and, just as much, the validation that separates a result that will hold from one that will not.
Because financial data has a low signal-to-noise ratio, few independent observations and patterns that change over time. A large network will fit noise unless the data is prepared and tested with great care. Every model in the course is compared with a simple baseline on data it has never seen.
Where is reinforcement learning used in trading?
Reinforcement learning trains an agent to make a sequence of decisions by trial and error, guided by rewards. Trading has many sequential decisions: how fast to sell a large position, where to quote, when to rebalance. That makes reinforcement learning for trading a natural fit in principle. In practice, its success depends on the quality of the simulator the agent learns in.
Its most credible applications are in execution and market making. In both, the problem is well defined, the reward is clear, and analytical models give a benchmark to beat. Applying it to choose which assets to buy, from historical prices alone, is much harder, because the agent can easily learn the quirks of a short history.
Who is this course for?
Quant researchers and data scientists working on systematic strategies who want research-grade methods.
What you will be able to do
- Build volume, dollar and imbalance bars and compare their statistical properties.
- Label events with the triple-barrier method and apply meta-labeling.
- Weight samples whose outcomes overlap in time.
- Apply purged and combinatorial cross-validation.
- Build and train feed-forward, recurrent, convolutional and transformer networks in PyTorch.
- Formulate a trading problem as a Markov decision process.
Syllabus, module by module
Why finance is different
Low signal-to-noise, non-stationarity, overlapping outcomes and few independent samples.
Sampling
Time, tick, volume and dollar bars, imbalance bars and event-based sampling.
Labeling
The triple-barrier method, meta-labeling and bet sizing from probabilities.
Neural networks from first principles
Layers, activations, back-propagation, optimizers and regularization.
Financial data for deep models
Scaling, sequences, labels, sample weights and the leaks that ruin results.
Sequential decisions
States, actions, rewards and why trading problems are sequential.
Markov decision processes
Bellman equations, value iteration and policy iteration.
What every course contains
Every paid course comes with the first seven and the certificate, whichever level it sits at. The proctored exam and the credential come when the course is taken as part of a Certificate, Diploma or Fellowship. Everything is self-paced: the material is yours to study in any order, at any hour, for the whole of the license.
- Workbooks. Written to train the thinking patterns and the important concepts of the course, with worked problems and the reasoning behind each answer, so that you learn to think the way the subject is thought.
- Audio edition. Focused, well-presented recordings with explainers and examples of the material, made to be interesting, relevant and entertaining, so that the general subject can be learned away from the desk.
- Notes. Pre-produced notes with every fact the course relies on, written out in full, so that nothing has to be transcribed from the films or the audio.
- Cheat sheet. One sheet with every important learning concept of the course on it, to keep beside you during the work and to revise from before the exam.
- Named sources. Every fact in the notes names its source, the paper, the book or the dataset, for anyone who wants to dig in further.
- Explainer films. Focused films, in the style of the sample films on the catalog page, that walk through the most difficult concepts of the course with visual explanations of a high standard.
- Transcript. A full transcript of every film and every audio recording, searchable and quotable.
- Proctored exam. Taken as part of a Certificate, Diploma or Fellowship, the course ends in a proctored multiple-choice questionnaire, with a pass mark of 60% and three retakes included. A course taken on its own is assessed by its coursework and has no exam.
- Verifiable certificate. A certificate of completion that a third party can verify: an employer sends the code on it to the credentials office and receives written confirmation of the course, the date and the result.
- Credential of the program. The Certificate, Diploma or Fellowship of the program, track or path the course belongs to, awarded when every course in it is passed. Which one depends on the program you choose.
How the certificate is earned
A written paper-style assessment with a computational component: you apply the methods of the course to a research question of your choice and write up the design, results and robustness checks as a paper. On a pass you receive the Course certificate in Financial Machine Learning, issued by the Institute with a verification code. A course taken as part of a Certificate, Diploma or Fellowship ends in a proctored online exam with a pass mark of 60%. The fee includes three retakes; each attempt after that carries a reduced proctored exam fee. A course taken on its own is assessed by its coursework and has no exam. Read how certification works.
Before you start
This course assumes the material of the courses below, or the same knowledge gained elsewhere.
This course counts toward
- Certificate in Machine Learning for TradingFive courses and a capstone on applying machine learning to markets without fooling yourself: labeling, features, purged validation, deep and language models, regimes and reinforcement learning.
- Fellowship in Computational FinanceThe top credential on the researcher path. Twenty-one courses and a dissertation supervised in writing and examined by two faculty members.
Questions applicants ask
Is this course based on Advances in Financial Machine Learning?
It covers many of the methods that Marcos López de Prado set out in that book and in related papers, alongside work by other authors. The book is recommended reading; the course is independent and not affiliated with the author or publisher.
What is meta-labeling?
A two-stage approach: a primary model or rule decides the side of a trade, and a secondary machine learning model decides whether to take it and how large to make it. It separates the direction decision from the sizing decision.
Does deep learning work for stock prediction?
Sometimes, and far less often than published results suggest. Market data is noisy and scarce compared with images or text, so deep models overfit easily. The course teaches how to tell a real gain from a lucky one.
Does reinforcement learning work for trading?
Its strongest results are in problems with a clear structure and a good simulator, such as optimal execution and market making. For predicting prices to pick trades, it faces the same data limits as other machine learning and adds problems of its own. The course covers both.