FPL Skill vs Luck: How Copilot Rank Measures Your Season Variance

Every FPL manager has thought it. “I should be ranked higher. The ball didn't drop for me.” Sometimes that's true. Sometimes it's an excuse. I built Copilot Rank to put a number on the difference.

How Much of FPL Is Luck?

Variance in FPL is huge. A captain returning a hat-trick versus a blank is a 24-point swing. A defender getting one bonus point or not is the difference between a clean sheet and a screen-throwing Saturday. Across 38 gameweeks, those swings stack into thousands of rank places either direction.

The cleanest way to measure it: replay your whole season scored on expected points (xPts) instead of actual points. Whatever rank falls out is the rank your picks deserved, holding luck constant. Compare that to your real rank. The gap is what the model attributes to variance.

What Copilot Rank Does

Copilot Rank shows two numbers side by side: your actual rank (FPL's overall rank) and your Copilot rank (the rank you'd have if every gameweek had been scored on xPts instead of actual points). Same picks, same captain, same chips, same hits. Different scoring engine.

The gap between the two is your variance signature. Below your real rank means you got unlucky and your picks deserved better. Above your real rank means the ball was kind to you.

How I Calculate It

• Replay every GW you played. I take the squad you submitted and re-score it on Copilot xPts. Captain ×2 (×3 with Triple Captain), bench ×0 (×1 with Bench Boost), Free Hit substitutions, autosubs, transfer hits, all preserved.
• Sum the season for your Copilot points total.
• Convert points to a rank using a curve sampled directly from FPL's overall leaderboard each week. ~20 (rank, points) pairs per GW, denser at the top where small point swings move rank a lot. Log-linear interpolation between tiers.
• Compare to your actual rank. The gap is what the model attributes to luck rather than skill.

For the luck percentile, I run the same replay across a stratified sample of ~6,500 managers (top 1k / 10k / 100k / 1M tiers). Each manager gets a delta = Copilot points minus actual points. The empirical distribution of those deltas is the field's luck distribution. Your delta lands somewhere on it. That's your percentile.

Change Your Luck With the Slider

The card has a draggable slider that runs from P0 (unluckiest) to P100 (luckiest). Drag the dot and the rank below it recomputes in real time. Same picks. Same captain. Same chips. Different luck level.

The slider answers questions managers actually ask themselves every season:

• If I'd had average luck (P50), what rank would my picks deserve?
• If the ball had bounced for me like it did for the top 10% of managers (P90), where would I sit?
• If I'd been as unlucky as the bottom 10% (P10), how far would I have dropped?

The math: at any percentile p, I look up the empirical luck delta at that quantile, subtract it from your Copilot points, and push that hypothetical points total through the rank curve. The result is your rank holding picks constant and varying only the luck. Useful for setting expectations: keep doing the same things, and at average luck this is roughly where you'll land.

Reading Your Luck Percentile

Your luck percentile is a single number from 0 to 100 telling you where you sit in the field this season. The slider on the card also lets you ask “what would my rank be at any other luck level?”

• P0 to P20: Unlucky to Very Unlucky. Your picks deserved a meaningfully better rank than you're sitting at. The model says variance cost you tens of thousands of places. Stay patient. The xPts were right; the bounces weren't.
• P40 to P60: On Par. Your rank matches what your picks should have produced. No big variance story either way.
• P80 to P100: Lucky to Very Lucky. Your rank flatters your picks. The ball carried you. Worth checking whether the same picks would still rank similarly without the lucky breaks. If not, expect regression.

Does Copilot Rank Actually Work?

Fair question. A metric that claims to separate luck from skill is only useful if it lines up with reality. Here's the cross- check I run, over the full 12,500+ manager lookup.

If Copilot rank were noise, the top-by-Copilot cohort would be randomly distributed across actual ranks. It isn't. The median actual rank inside each Copilot tier collapses tightly:

• Top 1k Copilot rank → median actual rank: 747
• Top 10k Copilot rank → median actual rank: 2,789
• Top 100k Copilot rank → median actual rank: 30,900

Each tier's median actual rank lands roughly inside that tier. About half of “best decisions” managers also finish inside the top 1k. Half the “top-10k decisions” cohort finishes top 3k. The picks captured by xPts replay correlate tightly with the rank managers actually achieve.

Where the two diverge is the variance signature. A manager at top-1k Copilot rank but 50,000th actual rank is a real signal: the picks were elite, the ball didn't drop. A manager at 100,000th Copilot rank but top-10k actual rank is the inverse: good rank, but the model says the picks weren't as good as the rank suggests. Both are meaningful patterns, not noise.

Why This Is Different From Your Overall Rank

FPL's overall rank tells you what happened. It does not tell you why. A manager at 150,000th could have been 50,000th with one more bonus point or 350,000th with one fewer. Overall rank bakes in every lucky bounce and every cruel deflection.

Copilot rank strips the bounces. It answers “if every player scored exactly the model's expectation, where would you sit?” The two numbers together are the full picture: what you got, and what you should have gotten.

Caveats

• It's not a season prediction. Past variance doesn't guarantee future regression. Some managers stay lucky for a season; some stay unlucky.
• It's not a verdict on your decisions. A captain pick can be the right call (highest xPts) and still blank. The model says the call was sound. The ball decides.
• xPts is a model, not ground truth. It is forward- looking and probabilistic. It under-weights sample-of-one events like specific tactical matchups, and systematically under/over-rates some archetypes (high-variance forwards, fixture-dependent defenders). Treat the gap as directional, not exact.
• The rank curve is sampled, not exhaustive. Roughly 20 (rank, points) tiers per gameweek with log-linear interpolation between them. ±5% rank accuracy on any single conversion is realistic, more in the long tail beyond top 1M.
• The 12,500+ manager lookup is a sample. It covers the leaderboard stratification (top 1k / 10k / 100k / 1M tiers) plus every team_id any user has queried via the warm-queue. That biases the sample toward engaged managers and toward people connected to FPL Copilot users. Stratifying mitigates but doesn't eliminate this. The full ~10M FPL universe behaves similarly in shape but the percentile thresholds may shift slightly with a fully random sample.
• Single-GW Copilot rank is noisy (~10% of luck comes from one big GW captain swing). Read the season-cumulative number, not the per-GW one, when evaluating skill.