A simple regression of T-bill yields on stablecoin flows would conflate the channel of interest with confounding co-movement. Aggregate stablecoin flows correlate with broad financial conditions: risk-on episodes draw funds into both cryptoassets and stablecoins while simultaneously affecting safe-asset yields, and Treasury yields themselves feed back into the opportunity cost of holding stablecoins. We therefore need a flow shifter that lives inside the stablecoin sector and is plausibly uncorrelated with broader macro and risk conditions.
We construct such an instrument following the granular instrumental variable (GIV) approach of Gabaix and Koijen (2024). The stablecoin sector is highly concentrated at the stablecoin-chain level: as Figure 3 shows, a single stablecoin-chain pair accounts for roughly one third of aggregate market capitalization, and the top three pairs together account for around 80%. In a market with a few large units, idiosyncratic shocks to those units do not cancel under cross-sectional aggregation but instead propagate into aggregate flows. These shocks are, by construction, orthogonal to common factors, including the risk-on co-movement and macro conditions that would violate the exclusion restriction. Therefore, the GIV approach is well suited to instrument aggregate flows.
For each unit i (a stablecoin-chain pair) on day t, let gi,t denote the growth rate of market capitalization, si,t−1 the lagged market-capitalization share, and Nt the number of active units. We define the size-weighted aggregate growth rate:
and the unit-level idiosyncratic residual:
The granular IV is the share-weighted cross-section of these residuals:
computed daily on the baseline stablecoin-chain panel described in Section 2.1.18
The instrument we use in our baseline specifications is the 5-day rolling sum:
which aligns with our 5-day flow measure; we also report results using the daily GIVt as a robustness check. Figure 4 plots both versions of the granular IV over the sample. Both series fluctuate around zero, exhibit no persistent trend, and pass standard stationarity tests.19
Table 2 reports the first-stage relationship between 5-day stablecoin flows and the granular IV. Column 1 shows the baseline: GIV(5d)t is significantly and positively correlated with aggregate 5-day flows, with a HAC-robust F-statistic of 54, well above conventional weak-instrument thresholds.20 A one-standard-deviation movement in GIV(5d)t corresponds to a 5-day stablecoin flow of approximately $0.6B. Column 2 reports the same regression using the daily granular IV, which is somewhat weaker (F = 30, flow per 1 SD = $0.3B) but is still comfortably above standard threshold levels. As a robustness check on the GIV construction, Column 3 uses a granular IV computed from DeFiLlama data alone, without the Token Terminal substitutions discussed in Section 2.1 (F = 42).
| (1) Baseline panel | (2) Baseline panel | (3) DeFiLlama-only panel | |
|---|---|---|---|
| Instrument used | GIV(5d)t | GIVt (daily) | GIV(5d)t |
| GIV coefficient | 1.14*** | 1.69*** | 1.04*** |
| (0.15) | (0.31) | (0.16) | |
| Constant | 0.64*** | 0.81*** | 0.68*** |
| (0.06) | (0.06) | (0.06) | |
| Observations | 1,362 | 1,362 | 1,362 |
| HAC-robust F-statistic | 54.4 | 30.2 | 42.0 |
| R² | 0.123 | 0.039 | 0.097 |
| 5-day flow per 1 SD of GIV | $0.61B | $0.34B | $0.54B |
Note: Dependent variable is the 5-day stablecoin flow Flow(5d)t in $B. The granular IV is expressed in percentage-point units of (weighted) cross-sectional growth; the reported coefficient is therefore the $B 5-day flow associated with a 1% movement in the granular IV. Columns 1 and 2 use the granular IV constructed on the baseline DeFiLlama–Token Terminal panel (5-day rolling sum and daily, respectively). Column 3 uses the granular IV constructed only from DeFiLlama data. Newey-West standard errors with one lag; *, **, *** denote significance at the 10%, 5%, and 1% level, respectively. HAC-robust F-statistic reported. Sample: January 2021 to March 2026.
The exclusion restriction is supported by both the construction of the granular IV and direct empirical tests of its correlation with potential confounders.
By construction, the granular IV is a share-weighted cross-section of unit-level idiosyncratic residuals after the size-weighted common factor has been removed. A principal-components decomposition of the underlying stablecoin-chain growth-rate panel confirms that the dominant common factor is modest: the first principal component explains only about 16% of cross-sectional variance, with the remaining variation supplying the cross-sectional dimension the granular IV exploits.
We complement this construction argument with a direct empirical correlation exercise. We compute the matched-frequency correlation of each version of the granular IV with twelve variables that span the principal channels through which broader risk sentiment, financial conditions, and sectoral demand could correlate with stablecoin growth: the VIX index, the Bloomberg US and Goldman financial conditions indices, the (log) S&P 500, the (log) gold and Brent prices, the (log) US dollar NEER, the (log) Bloomberg Galaxy Crypto Index, the (log) prices of Bitcoin and Ether, and sectoral demand from MMFs and short-term Treasury ETFs (the matched-frequency change in MMF assets under management and the matched-frequency sum of ETF flows, respectively). Frequency matching is essential: we correlate the 5-day-sum GIV(5d)t with 5-day changes in each confounder, and the daily GIVt with 1-day changes. Several of these variables also enter our regression as controls; the low correlations reported here are therefore informative on top of that absorption, since they rule out the granular IV inheriting common-factor variation through a different channel. The maximum absolute correlation across the twelve confounders is 0.10 for the 5-day-sum IV and 0.06 for the daily IV; the granular IV is in particular essentially uncorrelated with the VIX, the S&P 500, the FCIs, and the dollar — the principal channels through which risk sentiment could transmit between cryptoasset markets and Treasury markets. Table 6 in Appendix C.3 reports the full correlation matrix.
Finally, we ask where in the stablecoin-chain panel the identifying variation lives. For each trading day we demean the panel (i.e. within-day) and fit ĝi,t = αs(i),t + βc(i),t + ui,t by OLS with stablecoin and chain dummies, then accumulate sums of squares across days (Appendix C.4, Table 7). About 62% of cross-sectional dispersion in unit growth rates is common across stablecoins on a given chain on a given day, about 6% is common across chains of the same issuer, and the remaining 36% is genuinely pair-specific. The granular weights wi,t = si,t−1 − 1/Nt load on units that depart from both the issuer mean and the chain mean, so the identifying variation in the granular IV is dominated by the within-pair residual ui,t rather than by either fitted dimension. This property, by construction, sharpens the IV's orthogonality to issuer- or chain-level common factors.
18. We classify a unit as active at t if its market capitalization exceeds $1 million on both t − 1 and t, and require Nt ≥ 5 for GIVt to be defined. Growth rates are robust-winsorized at ±8 × 1.4826 × MAD to suppress occasional retroactive corrections in the raw chain-level data.
19. Augmented Dickey-Fuller tests reject a unit root at the 1% level for both versions.
20. See Stock and Yogo (2005) for the standard critical values and Montiel Olea and Pflueger (2013) for the equivalent thresholds in the presence of heteroskedastic and autocorrelated first-stage residuals.