Excel-Based Transparency Labs for Interpretability and Model Governance
Modern machine learning models are increasingly deployed within highly regulated financial environments, including anti-money laundering, fraud detection, sanctions screening, and broader risk management programs. While the computational sophistication of these models continues to advance, regulatory accountability remains unchanged: institutions must understand, govern, validate, and defend the systems they rely on.
The Excel-based workbooks linked on this page were developed to address a specific gap between model deployment and model comprehension. They are structured transparency labs built to expose the internal mechanics of machine learning algorithms step by step. The objective is pedagogical clarity and validation literacy rather than rapid deployment.
Each workbook reconstructs a machine learning method from first principles. Rather than calling optimized libraries, the logic of the model is expressed directly through formulas. Intermediate outputs, weights, updates, residual calculations, and aggregation mechanics are visible cell by cell. In this sense, Excel functions as a validation sandbox, every transformation can be traced, every computation examined, and every iterative improvement observed.
The featured implementation is designed to emphasize interpretability: gradient boosting, as well as attention-based architectures that illustrate core concepts behind modern AI systems. Across all implementations, emphasis is placed on interpretability, structural transparency, and alignment with model governance principles.
In regulatory environments governed by principles such as SR 11-7, conceptual soundness and effective challenge are not optional. Institutions must be able to articulate what a model does, how it behaves under different conditions, and why its outputs can be trusted. Black-box reliance without comprehension creates governance risk.
The purpose of these Excel transparency labs is therefore not to replace industrial-grade ML libraries, but to illuminate them. By reconstructing algorithms in a fully visible framework, practitioners can better understand the mechanics that underpin production systems.
These materials hosted here retain full authorship attribution to Chandrakant Maheshwari. While MRMIA does not guarantee workbook functionality, they are intended for educational and interpretive use, not production deployment. Users are encouraged to trace formulas, examine interpreted outputs, and engage thoughtfully with the workbook’s logic and assumptions.
Transparency is not an aesthetic preference in regulated finance. It is governance infrastructure.
Model Library
Each workbook is a step-by-step reconstruction of a machine learning algorithm — formulas visible, logic traceable.
This workbook reconstructs a Transformer encoder architecture with each step visible, explaining how attention distributes information across tokens and layers.
This workbook extends time-series forecasting logic into an attention mechanism, illustrating how temporal relevance is learned and used for prediction.
This workbook demonstrates ensemble training with decision trees, including bootstrap sampling and majority voting for stable classification behavior.
This workbook rebuilds boosting behavior with weighted learners, showing how misclassified examples influence subsequent models.
This workbook explores boosting behavior using logistic loss, making it possible to inspect the gradients, residuals, and iterative updates.
This workbook implements second-order reasoning by inspecting Hessian-inspired terms, showing how boosted trees can be optimized with richer curvature information.
This workbook illustrates panel regression mechanics using fixed effects and first differences, emphasizing how time and entity variation is handled.
