How to Analyze Backgammon Games: PR, EMG, and Cube Errors

Modern backgammon engines convert subjective sense («that move felt right») into a hard number — equity-loss. Combined with PR (performance rating), variation arrows, and separate cube-action scoring, this turns post-game review into a precise diagnostic. This guide walks through the metrics, what they mean, and three real timeline patterns you'll see in practice.

Why analyze your games at all?

Backgammon is a game of small edges. Most decisions have multiple plausible candidates, and the difference between them is fractions of equity — invisible without a tool that can rank them. Engine analysis gives that ranking. For each of your moves, the engine searches the candidate plays, computes equity for each, and scores your choice relative to the best one found. This produces:

• Per-move equity-loss — how much value you gave up vs the best move.
• Move grade — best / ok / doubtful / error / blunder, derived from the loss.
• Cube-action audit — separately for double offers and accept/refuse decisions.
• Aggregate PR — game-level performance index.

What is Performance Rating (PR)?

PR is the standard scalar quality metric. Defined as the average per-move equity-loss multiplied by 500 (the gnubg convention; XG uses a slightly different constant). The multiplier is arbitrary — what matters is consistent ranking.

PR	Level	What it means
< 5	Master	Engine-level play. Achievable on standard positions for very strong tournament players.
5–10	Expert	Few errors, no big blunders. Most tournament regulars sit here on average.
10–15	Strong club	Occasional error, rare blunder. Strong online play.
15–20	Intermediate	Multiple errors per game, an occasional blunder.
> 20	Improving	Frequent blunders. Reading the backgammon pillar first will pay back fast.

What is EMG: equity-loss per move?

EMG (Equity in Money-Game terms) is the per-move equity-loss expressed in normalized money-game units. Approximate grade boundaries:

• bestyour move equals the engine's top choice.
• okEMG < 0.02 (statistical noise).
• doubtfulEMG 0.02–0.08.
• errorEMG 0.08–0.20.
• blunderEMG > 0.20.

Exact thresholds are configurable per engine — gnubg, XG, and Nardex all calibrate slightly differently. The relative shape (errors are 4× doubtfuls, blunders are 2.5× errors) is universal.

Pattern 1: what does a clean game look like?

A well-played game looks like this on the timeline — the curve sits close to the midline, no spikes, only a couple of small doubtfuls:

Nothing dramatic to learn from such a game. The scope of analysis here is fine-grained: understanding why a doubtful was doubtful, often a subtle pip-count or builder question.

Pattern 2: what does a single hit-or-not blunder look like?

The most instructive blunders in backgammon are tactical decisions: should I hit this blot and accept the return shots, or play safe? When the engine disagrees with your choice by 0.40+, that's the position to study:

Open this position in the editor, look at the structure (is it pure race, contact, or holding game?), and check what the engine's preferred play was — usually you'll find a pattern that recurs in similar future positions.

Pattern 3: how do cube errors hide in clean positional play?

Two big spikes here aren't positional — they're cube blunders. The losing side missed a take on position 6 and missed a double on position 11. Positional play was clean, but the cube cost the match:

If your overall PR is high but you can't find any obvious positional blunder, check the cube column. Missed takes are the most common — beginners drop too eagerly under positional pressure that's actually take-able. See the doubling cube article for the math behind take/drop thresholds.

Why are cube errors scored as a separate dimension?

Engines score four cube-action types independently of move quality:

• Missed double — engine says you should have doubled, you didn't.
• Bad double — you doubled, engine says it was premature; opponent benefits from taking.
• Missed take — engine says you should have taken, you dropped.
• Bad take — you took a position the engine says you should have dropped.

Cube blunders are usually larger than positional ones in equity terms because they compound through match-equity transformations. A late take in a money game might cost 0.05 equity; the same call in a 7-pt match might cost 0.25 match-winning chance.

Which mistake patterns are most common in backgammon?

• Hit-or-not at 70/30. Hitting opens you up; not hitting cedes tempo. Engine-level distinction is often subtle and you can pattern-train on these with rollouts.
• Premature prime breakup. Playing builders aggressively and breaking a 5-prime to do it. Often you should hold the prime even at temporary cost.
• Missed re-entry calculation. When the opponent has builders pointing at your home, the entry math (rolls that hit a closed point) needs to be done — not felt.
• Race-vs-contact misdiagnosis. Pure race PR is dominated by pip count; contact PR adds tempo and timing. Knowing which game phase you're in is the first decision.

Where to go next

• Backgammon rules pillar — the foundation.
• How the Doubling Cube Works — math of take/drop.
• «Как анализировать партию в нардах» — same content for long narde (in Russian).
• Developers — API for scripted analysis over many games.
• Nardex vs gnubg vs XG — engine differences in the numbers.

FAQ

What is PR in backgammon?

In backgammon, PR (Performance Rating) is the standard quality metric. Defined as average equity-loss per non-forced move, multiplied by 500 (gnubg convention). Lower is better. PR scales: < 5 master, 5–10 expert, 10–15 strong, > 15 improving.

What is EMG in backgammon?

In backgammon, EMG (Equity-Loss in Money-Game terms) is the equity you gave away on a single move, expressed in normalized money-game units. A blunder is roughly EMG ≥ 0.20; an error is 0.08–0.20; doubtful is 0.02–0.08; ok is < 0.02. EMG is what drives the move grade.

Why does my backgammon PR depend on which engine I use?

In backgammon, different engines have slightly different equity surfaces (neural-network architecture, training data, search depth). gnubg-2-ply, XG, and Nardex will agree on big calls but differ at the third decimal. Don't compare PR numbers from different engines as if they're the same scale; use one engine consistently.

How do cube errors interact with positional PR in backgammon?

In backgammon, they're scored separately and shown in their own column. A position with clean positional PR can still have a high cube-PR if you missed takes or doubled too late. The «Reading Nardex analysis output» article covers how the API exposes both surfaces.

Can I run backgammon analysis programmatically over many games?

In backgammon — yes. The Nardex API exposes per-position analyze endpoints; you can score whole game logs by feeding each decision point. See developers/quickstart for the request shape and developers/comparison for how this differs from gnubg/XG.