AI Research COIN

COIN: volatility clustering vs direction persistence over the last ~3 years

COIN's intraday turbulence is measurably more predictable from day to day than its direction. Over 751 trading days spanning roughly three years, the normalized high–low range shows a positive lag‑1 Pearson r = 0.192 (p ≈ 1.1e‑7) and an OLS slope that explains about 3.7% of next‑day range variance; by contrast, daily returns have r = −0.038 (p = 0.2937), i.e., no detectable day‑to‑day persistence. The typical intraday move is large — mean normalized range ≈ 8.96% — so even a small autocorrelation matters for anticipating how much COIN might move, not which way.

Below you'll find the resampling, normalization, correlation and autocorrelation results, charts, and regression diagnostics that back this finding: volatility clusters; direction does not.

The research question

For COIN over the past ~3 years, is tomorrow's turbulence far more predictable than tomorrow's direction — is the daily high-low range strongly autocorrelated (calm begets calm, chaos begets chaos) while daily returns show essentially zero day-to-day persistence? Thesis: today's range predicts tomorrow's range with a robust positive correlation even as today's return has no bearing on tomorrow's, so you can forecast how much COIN moves but not which way — volatility clusters while direction stays a coin flip.

How this was measured

Resampled COIN minute bars to daily OHLCV. Defined 'turbulence' as normalized intraday range (high−low)/close to remove price-level effects. Computed next-day alignment for both range and close-to-close return (today vs tomorrow) and measured Pearson/Spearman correlations and an OLS slope for each. Also reported autocorrelation profiles at lags 1/5/10/20. Positive and significant lag-1 correlation in range alongside near-zero return autocorrelation would confirm volatility clustering with little directional persistence. Window covers the available daily span inside the ~3-year limit (2023-07-03 to 2026-06-30).

The key numbers

Trading days analyzed
751
Daily observations with non-null return and range
Pearson r: today range vs tomorrow range
0.1923
p-value (range → next-day range)
0.0000
Two-sided; p=0.0000 < 0.05 → detectable range autocorrelation
Pearson r: today return vs tomorrow return
-0.0384
p-value (return → next-day return)
0.2937
Two-sided; p=0.2937 ≥ 0.05 → no clear return autocorrelation
OLS slope (range_t → range_{t+1})
0.1923
Unitless; on normalized range
R² (range_t → range_{t+1})
0.0370
Explained variance for next-day range
Mean daily normalized range
8.9608%
Average (high−low)/close
Mean daily return
0.2149%
Return std (daily)
4.9866%

Reading the numbers

Across 751 trading days, today's normalized daily range predicts tomorrow's range with a small but statistically significant correlation (r = 0.1923, p ≈ 1.11e-7), while daily returns show essentially no day-to-day persistence (r = -0.0384, p ≈ 0.2937). The range regression explains only about 3.7% of next-day range variance (R² = 0.037), so predictability is real but modest.

The charts

Autocorrelation by lag — normalized range vs returns
What this chart says

This bar chart contrasts autocorrelation at several lags for normalized daily range versus close-to-close returns. Look at the tall bar at lag 1 for range (0.1923) and the smaller but still positive value at lag 5 (0.1683): range autocorrelation is strongest on the next day and wanes by 10–20 days (−0.0023 and −0.0118). By contrast, return autocorrelations sit near zero at all lags (lag 1 = −0.0384, lag 5 = 0.0096), supporting the idea that volatility clusters while direction does not persist.

Today vs tomorrow — normalized range
What this chart says

The scatter of today’s normalized range versus tomorrow’s (n = 750) shows a faint upward tilt consistent with the r ≈ 0.1923 and OLS slope ≈ 0.1923 reported in the stats. The cloud centers near the mean range (~0.0896 today, ~0.0895 tomorrow) with many points clustered there but extreme ranges up to 0.7545 visible as outliers. This means today’s range gives a real signal about how large tomorrow’s move is likely to be, but the relationship is loose (R² ≈ 0.037) so forecasts are noisy.

Today vs tomorrow — daily return
What this chart says

The scatter of today’s return against tomorrow’s (n = 750) is a diffuse blob with no clear slope, matching the near-zero correlation (r = −0.0384) and an insignificant p-value (≈0.2937). Returns have a small mean (~0.0022) and much larger day-to-day dispersion (std ≈ 0.0499), so yesterday’s sign gives no reliable guide to tomorrow’s sign. In short, direction looks like a coin flip while magnitude shows modest persistence.

Autocorrelation profile by lag

lag_daysrange_autocorrreturn_autocorr
10.1923-0.0384
50.16830.0096
10-0.0023-0.0067
20-0.01180.0132

Lag-1 regression summary (today → tomorrow)

modelslopepearson_rr2n_pairs
Tomorrow range vs today range0.19230.19230.037750
Tomorrow return vs today return-0.0383-0.03840.0015750

The takeaway

Yes — COIN's intraday turbulence is noticeably more predictable day-to-day than its direction. The normalized high–low range shows a positive lag-1 Pearson r = +0.192 (p = 1.109e-07) over 751 trading days, and the simple today→tomorrow regression has R² ≈ 0.037, meaning today's range explains about 3.7% of tomorrow's. By contrast daily returns have r = -0.038 (p = 0.2937), i.e., no detectable day-to-day return persistence. Statistically this is a real signal for volatility clustering (the range result is highly unlikely to be pure luck) but its practical size is small — a useful lean, not a tight forecast. Practically: expect calm days to slightly raise the odds of calm next days and big-range days to cluster, but don't count on predicting whether COIN will go up or down tomorrow.

The fine print