Editorial catalog
Articles on explainable AI in investment decisions
Free, informational articles on the practice of making AI reasoning in investment decisions explainable — and on the points where that explainability quietly breaks. Read online; inquire by email when you want a closer look.
What SHAP actually tells you about an investment prediction
A close reading of SHapley Additive exPlanations on a multi-asset model: what the per-feature scores mean, how correlated inputs distort them, and when a confident attribution is still a misleading one.
Attribution · SHAP · Correlated features
Model cards for investment systems, line by line
How to read a model card for a model that trades. We walk through training data, intended use, out-of-scope use and subgroup performance, and flag the disclosures investment models most often omit.
Model cards · Disclosure · Governance
Interpretable surrogates and the limits of approximation
When a shallow tree mimics a deep ensemble well, the complex model carried little extra signal. When it cannot, the gap is the interesting part. A guide to partial dependence and surrogate fidelity.
Surrogates · Partial dependence · Fidelity
Explanation drift after a regime change
Explanations that held in one market regime often degrade in another. We look at out-of-sample decay, attribution stability across volatility regimes, and what to re-check when the world moves.
Regime change · Out-of-sample · Drift
Counterfactual explanations for a portfolio tilt
What is the smallest change to inputs that would flip a model's recommendation? Counterfactuals are readable by non-technical owners, and they expose which inputs the model is actually sensitive to.
Counterfactuals · Sensitivity · Readability
Where interpretability claims quietly stop being honest
A checklist of the common gaps between an explainability claim and the evidence behind it: post-hoc tools presented as if they were the model, local explanations stretched to global conclusions, and more.
Honesty · Checklist · Limits
Every article is written to be read, not to close a sale. If a piece raises a question, the right next step is an inquiry — not a subscription.