Courses

Machine Learning for Quantitative Research

AdvancedPremium8 sections · 27 lessons · 81 questionsLog in to track progress

Several systematic funds test machine learning in their researcher interviews, and some run a separate round for it. The round does not test whether you can call a library. It tests whether you know what a model assumes, why it fails, and how you would know that it had failed.

Financial data makes those questions hard. A daily return is mostly noise, the relationships drift, and the number of independent observations is much smaller than the row count. Methods that work well on images or text often fit noise here. A model that explains one per cent of the variance of returns out of sample can be worth a lot of money.

This course teaches machine learning with that problem in mind. It starts with what a model learns and why it overfits, derives the bias-variance decomposition, and shows how gradient descent fits a model. It then covers linear models: ridge, lasso and elastic net as penalties and as priors, logistic regression with the cross-entropy loss, and support vector machines with kernels. A section on evaluation follows: cross-validation and the choice of k, classification metrics and calibration, forecast metrics for returns such as the information coefficient, and hyperparameter tuning that does not fool you.

The second half covers tree ensembles, what feature importance does and does not mean, and how to combine several forecasts. It then turns to the data: how to build features, label returns, set cross-sectional targets, handle missing values and outliers, and retrain when the market drifts. PCA and clustering follow, then neural networks and sequence models. The last section puts it together: how to choose a model family on noisy data, and the machine learning round itself.

Each lesson starts from an interview question, works the numbers by hand, and ends with coding problems that you write from scratch in Python. The research validation course covers leakage, cross-validation on time-ordered data and multiple testing in full. This course covers those topics only as far as a model needs them, and links to that course for the rest.

What a Model LearnsSupervised learning as minimising a loss: expected and empirical risk, what each loss predicts, and what a low signal-to-noise ratio changes.Free previewOverfitting and the Train-Test GapWhat overfitting is, how to diagnose it from two numbers, why noisy data makes it the default, and a benchmark that catches a pipeline fitting noise.Free previewThe Bias-Variance DecompositionThe derivation step by step, what each term does as a model gains capacity, and a case where shrinking an estimate cuts its error by more than half.Free previewGradient Descent and Its PathologiesThe least-squares update derived by hand, the learning rate that decides convergence, why scaling matters, and what stochastic gradients trade.
Ridge, Lasso and Elastic NetPenalties as constraints and as priors: what ridge shrinks, why lasso gives exact zeros, and why a lasso's chosen features are not the true ones.Logistic Regression and Cross-EntropyThe model, the loss it comes from, the gradient derived in two lines, why squared error trains badly on probabilities, and what imbalance does.Support Vector Machines and KernelsHinge loss against logistic loss, the maximum-margin classifier, the C parameter, the kernel trick and its cost, and why SVMs output scores.
Cross-Validation and the Choice of kHold-out against k-fold, the choice of k as a bias-variance trade-off, the leave-one-out shortcut, and what a cross-validated error measures.Classification Metrics and CalibrationPrecision and recall, ROC and precision-recall curves, log loss and the Brier score, and why a probability that sets a size must be calibrated.Forecast Metrics for ReturnsOut-of-sample R² against the right benchmark, the information coefficient and its t-statistic, why hit rate misleads, and breadth.Tuning Hyperparameters Without Fooling YourselfValidation and test roles, nested cross-validation, why the best of many scores is optimistic, random search and the one-standard-error rule.
Decision Trees and BaggingHow a tree chooses a split, why bagging and random forests cut a deep tree's variance, the floor that correlation sets, and out-of-bag error.Boosting and Gradient Boosted TreesTrees fitted in sequence to the errors left so far: one round by hand, boosting as gradient descent, leaf penalties, early stopping and bagging.Feature Importance and Its TrapsWhat an importance ranking shows: impurity against permutation importance, correlated features, partial dependence, SHAP values and noise columns.Combining Forecasts and StackingWhen averaging forecasts helps, why equal weights often beat fitted ones, stacking on out-of-fold predictions, and how to combine alpha signals.
Building and Selecting FeaturesTurning raw data into features: point-in-time construction, scaling across the universe, counting every feature tried, and selecting without leaks.Labelling Financial DataTurning prices into targets: fixed-horizon labels, volatility-scaled thresholds, the triple-barrier method, meta-labelling and overlapping labels.Cross-Sectional Targets and NeutralisationChoosing the target for a stock-selection model: market-relative and residual returns, sector neutralisation, rank targets, and neutral features.Missing Values and OutliersWhy financial data goes missing, imputing inside the training folds, the look-ahead in forward-fill, and outliers: winsorising and robust scaling.Non-Stationarity and RetrainingModels trained on changing markets: covariate shift against concept drift, rolling and expanding windows, exponential weighting and live monitoring.
PCA and Dimensionality ReductionPCA on a returns panel as a statistical factor model: eigenportfolios, why the later components drift, and the noise edge once the market is removed.Clusteringk-means and hierarchical clustering of returns: k-means++, the correlation distance, linkage, and testing whether clusters of assets are stable.
Neural Networks in BriefWhat a neural network computes, how backpropagation finds its gradient, and why its capacity is dangerous on a few thousand noisy observations.Training Neural Networks in PracticeInitialisation scales, Adam and AdamW, schedules, normalisation and dropout, and a debugging routine that starts from the loss at step zero.Sequence Models and AttentionVanishing gradients in recurrent networks, LSTM gates, attention and its square-root scaling, causal masks, and when a sequence model fits the data.
Choosing a Model Family on Noisy DataLinear models, tree ensembles or neural networks: what each assumes, what a 52% hit rate is worth, and how to choose without fooling yourself.The Machine Learning RoundWho runs a machine learning round, its four formats, the derivations to have ready, and how to answer a model design question.

Keep reading Machine Learning for Quantitative Research

27 lessons in this course, and every other premium course, on one subscription.

Subscribe now
  • Every lesson in every course, with the worked examples and interactive simulators
  • Graded questions on every lesson, with explanations for the wrong answers as well as the right one
  • The trainers, timed assessments and brainteaser library that go with them