DeFi pools advertise a yield. Almost nobody looks at what that number does next. This follows 7,853 of them from the day each crossed 20% APY to ninety days later.
Figures computed 26 August 2026 and fixed at that date — later builds do not recompute them.
Of 7,853 pools advertising 20% APY or higher at some point since February 2022, the median retained 33.5% of that rate ninety days on. 61.3% lost more than half. Over a quarter kept less than a tenth.
Treated as exponential decay, that is a half-life of roughly 57 days — the advertised number loses half its value about eight weeks after you read it. That figure assumes a constant rate of decay, which real pools do not obey. It is a way to hold the result in your head, not a measurement.
Retention varies by more than an order of magnitude between protocols. A uniswap-v2 pool keeps 6.0% of its advertised rate across 886 pools; an aerodrome-slipstream pool keeps 69.5% across 220. Sample sizes are shown because they are the only thing that makes a row worth reading.
| Protocol | APY kept | TVL left | Pools |
|---|---|---|---|
| gmtrade | 131.7% | 98% | 50 |
| defichain-dex | 71.7% | 74% | 40 |
| euler-v2 | 69.8% | 83% | 60 |
| aerodrome-slipstream | 69.5% | 70% | 220 |
| aave-v3 | 68.9% | 153% | 56 |
| convex-finance | 64.9% | 137% | 150 |
| sushiswap-v3 | 60.9% | 90% | 36 |
| curve-dex | 31.9% | 118% | 329 |
| orca-dex | 30.4% | 100% | 294 |
| raydium-amm | 24.8% | 83% | 699 |
| uniswap-v3 | 24.7% | 89% | 1,210 |
| balancer-v2 | 22.3% | 82% | 60 |
| kamino-liquidity | 19.8% | 110% | 77 |
| uniswap-v4 | 6.8% | 87% | 530 |
| uniswap-v2 | 6.0% | 65% | 886 |
| camelot-v2 | 1.9% | 81% | 35 |
gmtrade above 100% is not an error. Its pools' advertised yields genuinely rose across the period while total value locked held steady at 98% — growing fee revenue on stable capital, not the arithmetic illusion of a pool being drained. TVL is shown for every row so that distinction is checkable rather than asserted.
Advertised rates for every pool tracked by DefiLlama, one point per pool per day — 7.76 million observations across 16,199 pools, reaching back to February 2022. That history is backfilled from DefiLlama's own public per-pool charts, not recorded by us; Candor Labs' own six-hourly record begins 20 August 2026 and is not what this report measures. For each pool that ever advertised 20% or more, we take the median advertised rate over the seven days from that first crossing, and the median over the seven days beginning ninety days later. The ratio is what the pool retained.
Why seven-day windows
An earlier version of this analysis compared single days and reported a median retention of 14.9% — less than half the real figure. Picking the first day a noisy series crosses a threshold preferentially selects spikes, and comparing a spike against a normal value later exaggerates the decay. The windowed figure of 33.5% is the correct one. We mention it because the error inflated the finding, and errors that inflate findings are the ones that survive review.