Does one issuer's supply move before another's? Rolling F-tests for predictive precedence, in both directions, across four pairs.
As of September 8, 2026, 0 of the 8 directed tests across four stablecoin pairs cleared the 5% threshold over a 60-day window. The strongest was USDC → USDT, with an F-statistic of 2.57 (p = 0.114). No relationship cleared the threshold in all three windows. Granger causality measures predictive precedence within the sample, not economic causation.
Granger causality tests whether one stablecoin's market capitalization changes predict another's, beyond the second coin's own history (Granger, 1969). It is not classical causation: it is predictive precedence in the time-series sense, and a series can precede another simply by responding to a common factor sooner. The four pairs span the major stablecoin design types, and each is tested in both directions, because precedence is not symmetric.
| Direction | Pair | F | p | Result |
|---|---|---|---|---|
| USDC → USDT | USDT ↔ USDC | 2.96 | 0.097 | not significant |
| USDT → DAI | USDT ↔ DAI | 2.90 | 0.100 | not significant |
| USDC → DAI | USDC ↔ DAI | 2.80 | 0.106 | not significant |
| DAI → USDT | USDT ↔ DAI | 1.39 | 0.250 | not significant |
| USDE → USDT | USDT ↔ USDE | 0.94 | 0.341 | not significant |
| USDT → USDE | USDT ↔ USDE | 0.39 | 0.536 | not significant |
| USDT → USDC | USDT ↔ USDC | 0.13 | 0.718 | not significant |
| DAI → USDC | USDC ↔ DAI | 0.01 | 0.944 | not significant |
Each panel tests one direction at the 5% significance level, with 8 tests shown together. Even if none of the relationships were real, there would still be roughly a 34% chance that at least one panel appeared significant by chance alone.
A significant reading in a single window is therefore weak evidence on its own; the same directed relationship persisting across windows is the standard a result here should meet. No multiple-testing correction is applied or implied.
| Direction | Pair | F | p | Result |
|---|---|---|---|---|
| USDC → USDT | USDT ↔ USDC | 2.57 | 0.114 | not significant |
| USDT → DAI | USDT ↔ DAI | 1.81 | 0.184 | not significant |
| USDC → DAI | USDC ↔ DAI | 1.36 | 0.248 | not significant |
| USDT → USDC | USDT ↔ USDC | 0.38 | 0.539 | not significant |
| USDE → USDT | USDT ↔ USDE | 0.23 | 0.633 | not significant |
| DAI → USDC | USDC ↔ DAI | 0.03 | 0.856 | not significant |
| DAI → USDT | USDT ↔ DAI | 0.01 | 0.925 | not significant |
| USDT → USDE | USDT ↔ USDE | 0.00 | 0.964 | not significant |
Each panel tests one direction at the 5% significance level, with 8 tests shown together. Even if none of the relationships were real, there would still be roughly a 34% chance that at least one panel appeared significant by chance alone.
A significant reading in a single window is therefore weak evidence on its own; the same directed relationship persisting across windows is the standard a result here should meet. No multiple-testing correction is applied or implied.
| Direction | Pair | F | p | Result |
|---|---|---|---|---|
| DAI → USDC | USDC ↔ DAI | 4.81 | 0.031 | p < 0.05 |
| USDT → USDC | USDT ↔ USDC | 2.81 | 0.098 | not significant |
| USDC → USDT | USDT ↔ USDC | 2.44 | 0.122 | not significant |
| USDC → DAI | USDC ↔ DAI | 0.13 | 0.723 | not significant |
| USDE → USDT | USDT ↔ USDE | 0.13 | 0.723 | not significant |
| USDT → USDE | USDT ↔ USDE | 0.02 | 0.885 | not significant |
| DAI → USDT | USDT ↔ DAI | 0.00 | 0.956 | not significant |
| USDT → DAI | USDT ↔ DAI | 0.00 | 0.964 | not significant |
Each panel tests one direction at the 5% significance level, with 8 tests shown together. Even if none of the relationships were real, there would still be roughly a 34% chance that at least one panel appeared significant by chance alone.
A significant reading in a single window is therefore weak evidence on its own; the same directed relationship persisting across windows is the standard a result here should meet. No multiple-testing correction is applied or implied.
No directed test clears the 5% threshold in all three windows. Any relationship visible at one window and absent at another has held for less than the longer window covers.
Each window is a separate test on overlapping data.
Test. Granger causality F-test (Granger, 1969) with 1 lag, comparing a restricted autoregressive model in the target series to one that also includes the other series’ lag. The Frisch–Waugh–Lovell theorem is used for numerical stability: the coefficient on the other series is estimated from residuals after partialling out the target’s own lag from both.
F = ((RSSr − RSSu) ÷ q) ÷ (RSSu ÷ (n − k))
Thresholds. Critical values are computed for the residual degrees of freedom at each window rather than the asymptotic F(1, ∞), which is why they differ between windows: 30-day, 4.21 at 5% and 7.68 at 1%; 60-day, 4.01 at 5% and 7.10 at 1%; 90-day, 3.95 at 5% and 6.94 at 1%. Each test also carries its exact p-value.
Data. Daily market capitalization percent changes, which for dollar-pegged coins represent net minting and redemption. Levels are not used: they trend, and a Granger test on trending series produces spurious relationships. At least 20 usable observations are required, so a window with too many gaps is absent rather than estimated.
What the result is not. Granger causality measures predictive precedence within the sample. It is not evidence of economic causation, and it does not identify a mechanism: two series can show precedence because one responds to a common factor sooner than the other, with no relationship between the issuers at all.
Updated daily at 15:40 UTC. See the methodology for data sources and coverage.
Granger causality tests whether one time series provides statistically significant predictive information about another, beyond the series’ own history (Granger, 1969). For stablecoins, “USDT Granger-causes DAI” means that knowing yesterday’s USDT market cap change helps predict today’s DAI change.
USDT and USDC are both large fiat-backed dollars and tend to be highly synchronized: they often respond to the same macro signals at the same time, producing near-zero F-statistics. Testing additional pairs spans different stablecoin design types, which can have genuinely different response speeds and timing. Lead-lag relationships are more likely to appear across design types than within them.
No. Granger causality is predictive precedence, not classical causation. “USDT Granger-causes USDC” means USDT’s past values help predict USDC’s future values, and nothing more. Two series can show predictive precedence because one responds to a common factor slightly sooner than the other, with no relationship between the issuers at all.
The F-statistic measures how much better the unrestricted model, using both the series’ own lag and the other series’ lag, predicts the series than the restricted model using only its own lag.
The threshold depends on the sample size, so it differs by window: an F above 4.01 is significant at the 5% level over 60 days, and above 7.10 at the 1% level. Each test on this page carries its own exact p-value.
Granger causality tests assume the series are stationary. Raw market cap levels trend upward over time, which produces spurious statistical relationships. Daily percentage changes are approximately stationary, and for dollar-pegged coins they represent net minting and redemption flow.
One lag tests whether information from one issuer reaches the other within 24 hours, the granularity of the daily data. Adding lags can improve statistical power but risks overfitting on a window of this length. Whether a one-lag relationship holds across the 30-, 60- and 90-day windows is shown directly in the table above.