Case study · Investment decisioning

ML+AI-empowered driver trees

Know not just the valuation — but why, and how much to trust each assumption. The biggest win: AI and machine learning combine to give a powerful, evidence-backed answer to the question behind every deal — which investment, where, generates the best ROI?

−40%

committee review time per deal

100%

drivers traceable to source — auditable

Top 3

highest-leverage risks flagged upfront

key business indicator financial indicator value driver — controllable external — not controllable ƒ / + how drivers combine

high confidence medium low ▢ model-flagged (high leverage / uncertainty) link colour = estimated impact on value (low → high)

Boxes are typed (VDMN): the black key business indicator breaks down into financial indicators, then into the value drivers you can control and the external factors you can't. ƒ / + show how children combine; red links carry the most of the valuation; ▢ marks where a small change moves the answer most.

Research-led, not hand-drawn

We don't sketch driver trees ad-hoc. The structure here follows the Value Driver Modelling Notation (VDMN) — a peer-reviewed graphical notation for value driver trees distilled from 161 published trees (Matthies, 2025). It's why every box is typed — key business indicator, financial indicator, controllable value driver, or external factor — why the maths is explicit (ƒ / +), and why the figures reconcile at every level.

That discipline is what makes a tree simulatable — the precondition for the per-metric models, Shapley attribution and Monte Carlo search below. And the value-driver-versus-external distinction is decisive for investment: you can only put money behind the drivers you actually control.

Reference: B. Matthies, “VDMN: A Graphical Notation for Modelling Value Driver Trees,” arXiv:2512.14740 (2025).

From a tree to a decision engine

A tree that explains the number. The first driver tree did more than split enterprise value into parts — it gave a root-cause view of how each driver rolls up into the top metric. Building it also enforced something valuable on its own: every metric had to reconcile at every level, so the figures add up cleanly from the leaves all the way to the top.

One model per metric

We then stood up a machine-learning (ML) model factory — AI generating and tuning a dedicated model for each metric in the tree. With an ML model behind every node, Shapley curves gave a deep, quantitative read on each metric and on how all the drivers interact — including the non-obvious cases a static tree hides, where two drivers sitting in separate branches are actually negatively correlated, pulling against one another. The waterfalls below show that attribution, driver by driver.

What moves each driver — Shapley waterfalls

One model per driver, explained. Each plot breaks the model's prediction into the features that push the metric up (red) or pull it down (blue), from the expected baseline E[f] to the predicted value f(x). Several features are the metrics one level down the tree; others capture forces a tree alone can't show.

increases the metric decreases the metric

Search the whole problem space

With an ML model behind every metric and the drivers' interactions mapped, the last step is the one the committee actually cares about. Feeding the per-metric models into Monte Carlo simulation, the AI explores the full space of scenarios — every plausible combination of moves — instead of a handful of hand-picked cases. Because the value drivers are the levers you can pull and the external factors are the ones you can't, the search stays honest.

The biggest win: a powerful, evidence-backed answer to “which investment, where, generates the best ROI?”

AI and ML turn a static valuation into a decision engine — aiming each unit of capital where it returns the most, with the reasoning the committee can audit.

Why you can trust it

Every driver traces to its source, carries an explicit confidence, and the model surfaces its own uncertainty. Auditable by design.