r=0.77 doesn't mean what you think: correlation levels vs returns

When the first version of the STI methodology shipped, the headline finding was "correlation r=0.77 between STI 500 and the CS2 player base, 2020-2026." That number got cited a lot. It's clean, it's institutional-sounding, and it suggests the skin market tracks the health of the underlying game.

This post is the honest unpacking. What that number actually means, and what it doesn't.

TL;DR

• r=0.82 in levels (raw index value vs raw player count) — measures long-term trajectory alignment
• r=0.26 in monthly returns (month-over-month changes) — measures real co-movement, statistically defensible
• r ≈ 0 over the trailing 12 months — skin market has developed its own dynamics post-CS2 launch
• All three are true. They mean different things, and mixing them up is how analysts get burned.

The trending-series problem

When two series both grow monotonically over time, correlation between them tends to look high. This is the classic spurious correlation trap: two variables that rise together don't require a causal link.

Famous example: the divorce rate in Maine and US per-capita margarine consumption have r=0.99 across the same decade. Are they causally related? Of course not. They just both happen to trend in the same direction during the period studied.

The CS2 skin market grew a lot between 2020 and 2026. The player base also grew a lot. Correlating their absolute values (levels) gives r = 0.82. Each 1% rise in player count corresponds — on average, over 6 years — to roughly 1% rise in the STI 500.

That's a real data point. It's not fraud. But does it imply causation? No. It's perfectly compatible with both:

1. The skin market grew for different reasons (CS2 launch hype, crypto bull market, scarcer cases) AND the player base grew for partially separate reasons. Both happened in the same window without one causing the other.
2. Both are functions of a deeper variable (gaming engagement as an emerging asset class, macro risk appetite, low-rate environment in 2020-2021). Causal — but not directly between the two series themselves.

What returns correlation actually shows

To filter out the trend and measure genuine co-movement, you use returns correlation — how much STI 500 changes in a month when player count changes in the same month. Rates of change, not absolute values.

On the same dataset, that correlation drops to r = 0.26.

In plain English:

• If player count rises 5% in a month, STI 500 rises on average ~1.3% in the same month (5 × 0.26).
• High variance around that average — other factors dominate in any specific month.
• Player count is not a strong predictor of month-to-month changes in the skin market.

A real relationship — but much weaker than the r=0.82 levels number suggests.

The trailing 12 months

Restrict the window to the trailing 12 months only (rolling). The correlation drops further.

r ≈ 0

Effectively no relationship between monthly changes in player count and monthly changes in the skin market over the most recent year.

Compatible hypotheses

1. CS2 maturity. Counter-Strike 2 launched in late 2023. By 2024, the player base had stabilized at roughly 1M to 1.5M daily — monthly changes are now mostly seasonality noise, not real growth.
2. Skin market decoupled. New cases (Kilowatt, Fracture series, etc.), community drama cycles, drop-pool policy changes, and Genesis Terminal in 2025 created drivers independent of player count.
3. Asset class maturing. CS2 skins have become an asset class with its own dynamics — analogous to how Bitcoin moved past "peer-to-peer cash" narrative and started picking up macro drivers (rates, dollar, risk appetite) that have nothing to do with the original use case.

Recent academic literature converges with this. The Frontiers in AI paper (Nov/2025) analyzing ~12,000 CS2 skins notes explicitly:

Skins develop dynamics linked to game lifecycle AND periodic updates, not direct player count mapping.

Our rolling-12m r ≈ 0 corroborates the literature.

Why we changed how we report it

The earlier version of the methodology surfaced only "r=0.77." The current version reports three numbers:

• Levels: r=0.82
• Returns: r=0.26
• Rolling 12mo: r ≈ 0

With explicit interpretation of each. Why the change?

Because a technical reviewer (quant, analyst, academic) spots this distinction in 10 minutes. Publishing only r=0.77 without the nuance reads as either naive or dishonest. Neither signals rigor. Publishing all three with context is a feature for the sophisticated reader, not a weakness.

What this means for you

If you allocate to skins for diversification

Good news: skins have their own dynamics. They're not purely a function of gaming engagement. Diversification within the crypto+alternatives bucket is more real than the player-count number alone suggests.

If you use STI as a market signal

Use player count as a directional macro indicator (rises with the market over the long term), not a monthly predictor. If player count drops 30% in a month, don't expect STI 500 to drop 30% — r=0.26 implies roughly 8% movement, and even that with high variance.

If you're a technical reviewer

The methodology computes 8 correlation configurations across the same data — levels (raw values), monthly returns, rolling 12-month windows, with and without lag, with and without de-trending. The full configuration matrix and audit notes are documented in the methodology page; implementation details available on request.

A note on editorial honesty

Traditional financial products tend to hide nuance like this. It's easier to sell "correlation r=0.77, reliable signal" than "correlation depends on how you measure, and trailing 12mo it's actually zero."

Editorial bet: in a market where most analysis is forum-post quality or recycled brokerage takes, methodological transparency is real differentiation. A serious technical reviewer respects a product that publishes three numbers with interpretation more than a product hiding behind a single polished one.

Disagreement, corrections, and challenges are welcome — contact direct via /about.

Technical appendix — the 8 configs we ran

For the curious or the skeptical, this is the output ofscripts/correlation-sensitivity.ts against the production backup backup-2026-04-23:

N = 72 months for levels, 71 returns for arithmetic.

Bottom line

A single correlation number is almost always less honest than three numbers with interpretation. r=0.82 in levels is real but partly trend artifact. r=0.26 in returns is the more rigorous co-movement measure. r ≈ 0 in the trailing 12mo is the recent regime — skins have decoupled.

When you read "r=0.77" about anything, ask: levels or returns? Whole sample or rolling? The answer changes what the number means.

Fernandes, J. (2026). r=0.77 doesn't mean what you think: correlation levels vs returns. Skin Trackers. https://skintrackers.com/en/blog/correlation-levels-vs-returns-cs2-skins

Jorgin Fernandes, "r=0.77 doesn't mean what you think: correlation levels vs returns," Skin Trackers, 2026, https://skintrackers.com/en/blog/correlation-levels-vs-returns-cs2-skins.

@misc{skintrackers_correlation-levels-vs-returns-cs2-skins,
  author = {Jorgin Fernandes},
  title  = {r=0.77 doesn't mean what you think: correlation levels vs returns},
  year   = {2026},
  url    = {https://skintrackers.com/en/blog/correlation-levels-vs-returns-cs2-skins},
  publisher = {Skin Trackers}
}

https://skintrackers.com/en/blog/correlation-levels-vs-returns-cs2-skins

Disclosure: Methodology choices are documented in /en/methodology. Critique welcome via /en/sobre contact. Analysis is editorial — not a buy or sell recommendation. Past correlations don't imply future ones.