stefan-jansen--machine-learning-for-trading/case_studies/README.md

Case Studies

Nine case studies thread through Chapters 6–20, applying the same ML4T workflow to different asset classes, frequencies, and trading constraints. Each one defines a universe, builds labels and features, trains models from linear baselines through deep learning, and evaluates strategies through backtesting, portfolio construction, cost analysis, and risk management.

What's in this release. This directory currently ships the shared utils/ package the chapter notebooks build on, plus each case study's canonical config/setup.yaml (universe, costs, walk-forward CV, labels). The per-case-study notebook pipelines, generated data, and cross-study results are released in book order over the coming weeks — watch or star the repo to follow along. Cross-case-study results are synthesized in Chapter 20.

Overview

#	Case Study	Asset Class	Frequency	Universe	Primary Label	What It Explores
1	ETFs	Multi-asset ETFs	Daily	100 ETFs	fwd_ret_21d	Cross-asset momentum and mean-reversion
2	Crypto Perps Funding	Crypto perpetual futures	8-hourly	19 pairs	fwd_ret_8h	Funding-rate arbitrage on perpetuals
3	NASDAQ-100 Microstructure	US equities (intraday)	15-min	114 stocks	fwd_ret_15m	Intraday microstructure signals from order flow and the LOB
4	S&P 500 Equity + Options	S&P 500 equities	Daily	634 stocks	fwd_ret_5d	Equity selection enhanced with implied-volatility features
5	US Firm Characteristics	US equities (fundamental)	Monthly	~2,500 stocks	fwd_ret_1m	Firm-level characteristics panel (size, value, momentum, quality)
6	FX Pairs	G10 currency pairs	Daily	20 pairs	fwd_ret_1d	Carry and momentum across major currency pairs
7	CME Futures	Multi-sector futures	Daily	30 products	fwd_ret_5d	Term-structure and roll-yield signals
8	S&P 500 Options	S&P 500 equity options	Daily	S&P 500 straddles	fwd_ret_dh_10d	Options-only strategies (straddles, delta-hedged positions)
9	US Equities Panel	Broad US equities	Daily	~3,200 stocks	fwd_ret_1d	Broad cross-section with classic factor exposures

Pipeline Stages

Each case study follows the same chapter progression. Notebooks are numbered sequentially, with each number mapping to a chapter:

Stage	Chapter	Typical Notebook	What It Produces
Feasibility	Ch6	01_feasibility_analysis	Universe and cost feasibility evidence for the canonical config/setup.yaml
Labels	Ch7	02_labels	Forward returns, walk-forward CV splits
Features	Ch8	03_financial_features	Momentum, volatility, carry, and domain-specific features
Temporal	Ch9	04_model_based_features	ARIMA, HMM, spectral features from walk-forward fits
Evaluation	Ch7–9	05_evaluation	Feature-label IC diagnostics
Linear	Ch11	06_linear	Ridge, LASSO, ElasticNet baseline predictions
GBM	Ch12	07_gbm	LightGBM predictions with Optuna
Tabular DL	Ch12	08_tabular_dl	TabM / neural tabular predictions
Deep Learning	Ch13	09-10_dl_*	LSTM, TCN, TSMixer, PatchTST, N-BEATS predictions
Latent Factors	Ch14	*_latent_factors, *_pca, *_ipca, *_sdf	PCA, IPCA, CAE, SAE, SDF factor models
Causal	Ch15	*_causal_dml	Double ML treatment effect estimates
Backtest	Ch16	*_backtest	Strategy simulation results
Analysis	Ch16	*_backtest_analysis	Performance attribution and reporting
Portfolio	Ch17	*_portfolio_management	Allocation methods and portfolio construction
Costs	Ch18	*_costs	Transaction cost impact analysis
Risk	Ch19	*_risk_management	Drawdown controls, position limits, risk budgets
Synthesis	Ch20	*_synthesis	End-to-end strategy assessment

Not every case study has every model type — the exact notebook set depends on the dataset. Each case study's own README (shipped with its pipeline) documents its complete table.

Directory Layout

Once a case study's pipeline is released, it follows this structure:

case_studies/{id}/
├── README.md                 # Dataset profile, pipeline table
├── config/
│   └── setup.yaml            # SSOT: universe, costs, CV, labels   ← shipped now
├── 01_feasibility_analysis.py / .ipynb       # Numbered notebook sequence
├── 02_labels.py / .ipynb
├── ...
├── data/                     # Labels and features (gitignored, reproducible)
│   ├── labels/
│   └── features/
└── run_log/                  # Model registry + runs (gitignored)
    └── registry.db           # Content-addressed per-run artifacts

Reproducibility

• config/setup.yaml defines the trading setup, cost model, and evaluation protocol — the single source of truth for each case study.
• The run log (Chapter 6.7) records every model run, content-addressed by its config hash, in run_log/registry.db.
• Generated artifacts (data/, run_log/, strategy/) are gitignored but reproducible by running the notebook sequence from the repository root.