Interpretable Machine Learning · Finance
When models explain themselves, decisions improve.
A structured lecture series on making ML models in finance readable, auditable, and defensible - without sacrificing predictive performance.
Is interpretability actually worth the tradeoff?
Most practitioners come in with the same doubt: choosing explainability means accepting a weaker model. That assumption has been tested against real financial datasets, and the gap is smaller than most expect - often under 1–2% in AUC on credit scoring benchmarks.
The real cost isn't accuracy. It's the regulatory and operational risk of deploying a model that no compliance team can interrogate. SHAP-based explanations and monotone constraints don't replace performance - they make performance defensible.
This series addresses that hesitation directly, with worked examples from credit risk, fraud detection, and portfolio attribution - not theoretical proofs.
Grounded in practice, not just theory
The material draws from production ML systems at financial institutions - not academic toy problems. Methods taught here have been applied to real regulatory submissions and model governance reviews.
SHAP in depth
Kernel SHAP, Tree SHAP, and interaction values - applied to gradient boosted models on tabular financial data.
Regulatory alignment
Mapping model explanations to SR 11-7 and EBA guidelines - what auditors actually ask for and how to answer.
Model stress testing
Identifying fragile decision boundaries and documenting model behaviour under distributional shift.
Practitioners describe the moment they could finally explain a model decision to a senior risk officer - and have it hold up to scrutiny.
Seeing yourself in the material
Credit risk modeller, 3 years in
Builds XGBoost models that perform well internally but faces pushback at model committee because nobody can trace a prediction back to a feature. Needs a structured way to present SHAP outputs that non-technical stakeholders accept.
Quantitative analyst moving into ML
Comfortable with statistical inference but new to tree-based models. Wants to understand what interpretability frameworks actually guarantee - and where they fall short.
ML engineer supporting compliance
Responsible for model documentation ahead of an internal audit. Needs to produce explanation artefacts that satisfy reviewers without rebuilding the model from scratch.
What stays useful after the lectures end
Skills in model interpretability compound over time. Every new model you build, every governance review you face, draws on the same conceptual toolkit - and it transfers across domains.
A reusable explanation workflow
The process for generating, validating, and presenting SHAP explanations doesn't change when you switch models or datasets. You build it once and adapt it.
Vocabulary that crosses team boundaries
Being able to describe model behaviour in terms that risk officers, compliance teams, and engineers all recognise changes how decisions get made around a model.
Documented reasoning for model governance
The artefacts produced during the course - explanation summaries, stability checks, feature attribution tables - are directly usable in model risk documentation.
Clearer instincts about model failure
Understanding what a model explanation reveals about decision boundaries helps you spot brittleness earlier - before it surfaces in production or in a review.
Stated in their own words
Participants describe concrete changes in how they work - not general satisfaction. These are specific before-and-after accounts from people who completed the full series.
"Before this, I was producing SHAP plots that I couldn't fully explain to a non-technical audience. I knew the values were correct but I didn't have a coherent story around them. After the third module, I restructured how I present model outputs entirely - the committee accepted our credit model without a single revision request, which had never happened before."
SHAP outputs produced but not integrated into governance documentation
Explanation artefacts became a standard part of every model submission
"I came in knowing Python and gradient boosting but had no framework for explaining model decisions to auditors. The section on regulatory mapping was exactly what I needed - specific, not abstract."
"The stress-testing module changed how I think about model monitoring. I now run stability checks as a routine step, not just before a review."
"Interpretability had always felt like a concession. This series reframed it as a tool for understanding what a model is actually doing - which turned out to be more useful than the accuracy metric alone."