How Cyprus Apex rating works

Idea in one sentence

Strength is born at the apex (ITF, Tennis4All 1000/1500, high TopTennis) and flows through who beats whom — weighted by round, score, and freshness. A soft-pool winrate without that world cannot invent Elite.

How the number is calculated

1. Seed on strong matches. Playing ITF / T4A Pro–Elite / high TTT adds authority. Deeper rounds (QF/SF/F) and clearer scores add more. The loser still gets a small contact credit — showing up matters.
2. Propagate through the graph. Beating someone with authority transfers a chunk of it. Beating an isolated club player barely moves you. Club edges are weaker carriers than apex edges.
3. Last-N H2H. Only the freshest 3 matches per pair get full weight; older repeats are down-weighted (anti-farm).
4. Mild time decay so old apex glory fades (~30%/year).
5. Display scale: rating = 1000 + 180 · ln(1 + raw_authority). RD encodes apex connection: ≤150 means recent strong-event contact.
6. Elite gate: shelf Elite needs both a high display rating and RD ≤ 150. High score without recent apex caps at Advanced.

What we compared

Walk-forward on ~11k Cyprus singles. Glicko still wins raw next-match accuracy (~70% recent-12m, ~78% on strong slate). Apex wins fairness: soft-Elite drops from ~43% (Glicko μ≥1700 with no recent strong play) to 0% with the gate; deep-run players sit higher (~84th percentile vs ~79th).

So we ship Apex for the public ladder / levels. Glicko is still computed in the background for research and match-forecast work.

Are we validating wrong?

Judging Apex only by “did the favourite win the next club match?” asks the wrong question. Apex is built to answer: who has proven strength in the strong world, and who is connected to it? A model that refuses to crown soft-pool farmers will look “worse” on casual upsets — and that can be correct behaviour.

Better checks for Apex:

• Soft-Elite rate (high score, no recent apex) → should be ~0 on the Elite shelf
• Deep ITF/T4A runs sit high on the ladder
• Repeat H2H farm moves the score less
• Case audit of known tops vs inflated club specialists
• Accuracy on the strong slate separately from all matches

Where Apex “errors” cluster

• Cold / unconnected players — both near zero authority → coin-flip forecasts (expected).
• Club upsets — Apex underweights soft pools; Glicko may “call” them better while being less fair for Elite.
• Calibration — raw authority gaps are not tuned as probabilities; do not read Apex as a betting model yet.
• Sparse apex coverage — strong players who rarely enter ITF/T4A climb only via 1–2 hop wins over connected opponents.

How to read RD (Apex)