The Cadence Architecture — Post 2 — “The Random Forest”

The Cadence Architecture Post II  ·  Forensic System Architecture  ·  Sub Verbis · Vera

EVIDENTIARY BASIS: DOCUMENTED

The Random Forest

// 2019–2025 — the prediction model that's already been running in Tour de France broadcasts for years

---

---

Randy Gipe · Claude / Anthropic  ·  2026  ·  Trium Publishing House Limited  ·  Forensic System Architecture

Model Diagnostic — Post II

The prediction layer this series keeps circling back to isn't a future capability. It's been in production, in public, since before this series existed.

2019 — The Model

NTT builds a breakaway-success predictor for Tour de France broadcasts: a random forest model, roughly 35 features, re-run every 10 kilometers of the course.

Inputs

Gap size, terrain gradient, team composition in the break, rider history, and GC standings — none of it exotic, all of it public race data.

2025 — Formalized

A peer-reviewed paper integrates energy expenditure, aerodynamic drag, and crash probability into a single breakaway-timing optimization — moving the same idea from broadcast novelty to operations research.

Diffusion

Betting platforms are already running comparable predictive models off live GPS telemetry, recalibrating odds mid-race — the same technique, a different customer.

I  ·  The Model

---

Six years before this series existed, NTT built a machine learning model to answer a single question for Tour de France broadcasts: will today's breakaway survive to the finish? The model is a random forest — an ensemble of decision trees, not a single exotic algorithm — fed by roughly 35 features and re-run every 10 kilometers as the race unfolds. It's been quietly doing this since 2019.

None of the inputs are secret or proprietary. Gap size to the peloton, terrain gradient, how many teams have riders in the break, those riders' history in similar situations, and where the overall standings sit — all of it is public race data, the same information a knowledgeable fan watching the broadcast already has access to. The model's contribution isn't better data. It's doing the arithmetic on all of it, continuously, faster than a person could.

II  ·  The Formalization

---

What's changed since 2019 isn't the existence of the idea — it's how seriously the idea is now being treated. A 2025 academic paper takes the same basic question NTT's model answers for television and turns it into a formal optimization problem: given a rider's power output, the aerodynamic drag they face, and the accumulating risk of a crash, what's the mathematically optimal moment to attempt a breakaway? That's a meaningfully different level of rigor than a broadcast graphic — peer-reviewed, reproducible, and explicit about its assumptions in a way a TV predictor never has to be.

35

Features feeding NTT's breakaway model

Re-run every 10 kilometers of the course, in production for Tour de France broadcasts since 2019 — six years before anyone was writing about an "AI arms race" in cycling.

III  ·  The Diffusion

---

The same underlying technique — live recalibration of an outcome probability, fed by GPS telemetry as the race moves — is no longer confined to broadcasters. Betting platforms are running comparable predictive models off the same public race-tracking data, adjusting odds in real time as gaps open and close. It's the clearest available evidence that the tool isn't exotic or hard to build. It's diffused into an entirely different industry with different incentives, using the same public inputs.

Evidentiary Note Documented vs. Interpretation

NTT's model, its feature count, its refresh cycle, the 2025 optimization paper, and the betting industry's use of comparable live modeling are all independently documented. What is not documented anywhere in public reporting: whether any WorldTour team's own performance staff uses this specific tool, or an equivalent one, as an in-race tactical instrument. NTT built this for television audiences, not team radios. Post III's argument about teams adopting similar tools internally is our forward read, not a claim made here.

FSA Wall — Post II

NTT's 2019 breakaway-prediction model, its random forest architecture, feature count, and 10-kilometer refresh cycle are drawn from Cyclingnews's contemporaneous report on the tool, treated as Tier 1. The 2025 academic paper formalizing breakaway timing as an energy/drag/crash-risk optimization is drawn from its Royal Society Open Science publication, treated as Tier 1 peer-reviewed research. The diffusion of comparable predictive modeling into cycling betting markets is drawn from Pez Cycling News's reporting on machine learning in race wagering, treated as Tier 2. A supplementary account of AI prediction tools in cycling more broadly is drawn from ProCyclingUK commentary, treated as Tier 2 and used only for corroborating context, not as a primary claim source.

Up Next — Post III

The tool is real, it's public, and it isn't hard to build. Post III, The Second Order, is where this series stops reporting and starts arguing: what happens once every team is running some version of the same model.

The Cadence Architecture  ·  Series Navigation

MastheadWhere Pacing Meets Prediction

Post IThe Permitted Field

Post IIThe Random Forest

Post IIIThe Second Order

Post IVThe Camera's Pulse