ML fundamentals applied to finance

5.1.1 Why financial ML is different: low signal-to-noise ratio, non-stationarity, scarce and contested data.
5.1.2 Supervised, unsupervised and reinforcement learning; scikit-learn: pipelines, transformers, hyperparameter search.
5.1.3 Appropriate metrics, not just accuracy: economic metrics (signal Sharpe, P&L after costs).

Feature engineering

5.2.1 Price and volume features and microstructure features: multi-horizon returns, volatility, normalised indicators, imbalance, VPIN.
5.2.2 Point-in-time fundamentals, macro, sentiment, alternative data.
5.2.3 Fractional differencing to preserve memory while achieving stationarity.
5.2.4 Normalisation without leakage (rolling statistics), cross-sectional rank transform.
5.2.5 Feature selection and importance: MDI, MDA, SFI, SHAP; feature clustering.

5.3.1 Fixed-horizon labelling and its problems.
5.3.2 The triple-barrier method.
5.3.3 Meta-labeling: primary model and secondary model for sizing or filtering.
5.3.4 Sample uniqueness, weights by overlap and by return; sequential bootstrapping.
5.3.5 Event-based sampling (CUSUM filter).

Fig. — Triple-barrier method: profit-taking, stop and time limit (schematic)

5.4.1 Why random k-fold fails.
5.4.2 Purged K-Fold with embargo; CPCV.
5.4.3 Walk-forward with retraining; expanding versus rolling windows.
5.4.4 Backtest of the ML signal with costs; Deflated Sharpe over the experiments run.

5.5.1 Regularised regressions (Ridge, Lasso, Elastic Net); logistic.
5.5.2 Trees, Random Forest, bagging.
5.5.3 Gradient boosting: XGBoost, LightGBM, CatBoost, the workhorses for tabular data.
5.5.4 Hyperparameter tuning with Optuna, and the risk of overfitting the validation.
5.5.5 Unsupervised: PCA, clustering (regimes, asset grouping), autoencoders.
5.5.6 Cross-sectional ranking models (learning to rank) for stock selection.

5.6.1 Multi-backend Keras 3 (JAX, TensorFlow, PyTorch): configuration, functional and subclassing API.
5.6.2 MLPs for financial tabular data; regularisation (dropout, weight decay, early stopping).
5.6.3 1D CNNs and LSTM/GRU for sequences; windows, scaling and leakage.
5.6.4 Callbacks, custom losses (for example, Sharpe-based), export and inference.

5.7.1 Tensors, autograd, nn.Module, Dataset and DataLoader, training loop; GPU and MPS, torch.compile, Lightning.
5.7.2 Time series architectures: TCN, N-BEATS and N-HiTS, PatchTST, iTransformer; transformers with attention and patching.
5.7.3 Probabilistic models: quantile loss, DeepAR, confidence intervals.
5.7.4 Graph neural networks (GNNs) for relationships between assets (introduction).
5.7.5 End-to-end portfolio optimisation (differentiable layers, Sharpe loss).

5.8.1 What they are: massive pre-training and zero-shot forecasting.
5.8.2 Families: Chronos, Chronos-Bolt and Chronos-2, TimesFM-2.5, Moirai-2.0, TiRex, TimeGPT; Kronos, pre-trained on OHLCV.
5.8.3 Practical use: zero-shot inference, covariates, fine-tuning, probabilistic forecasting.
5.8.4 Evidence in finance: in a 2026 benchmark (arXiv:2606.27100) on the daily returns of AAPL, AMZN, GOOG, JPM and META, Moirai-2.0 and TimesFM-2.5 achieved the best average ranks (2.9 and 3.1), but their improvements over the random walk were “small and sparse”. Another study found that strategies built on these models perform considerably worse than specialised reference models. Lesson: useful for volatility, covariates and series with structure; scepticism about forecasting returns directly.
5.8.5 Lab: foundation model versus GARCH and LightGBM in forecasting realised volatility.

5.9.1 MDPs, Q-learning, policy gradient, PPO, SAC; Gymnasium environments and reward design with costs.
5.9.2 Realistic applications: optimal execution, market making, dynamic hedging.
5.9.3 Why RL tends to overfit in directional trading; FinRL as a critical reference.

5.10.1 SHAP, partial dependence, stability of importances.
5.10.2 Data drift and concept drift, regime detection.
5.10.3 Ensembles, signal combination and synthetic data (GANs and diffusion: state of the art and limits).