Archive / Crypto & DeFi / E878
Protocol fee growth showed no detectable next-day edge
Going long tokens whose protocol fees grew fastest and short those whose fees grew slowest earned +4.3 bp a day after costs over 846 test days, with a 95% interval from -9.7 to +18.3 bp, below the +8 bp required.
Public analytics dashboards show how much users pay each DeFi protocol in fees, while most perpetual-futures traders watch only price. The belief is that tokens of protocols with rising fees keep rising. We ranked about 229 tokens each day by fee growth (last 7 days against the prior 28), went long the top fifth and short the bottom fifth for 24 hours, and charged trading costs. Over 846 test days the pair made +4.3 bp a day net, with an interval from -9.7 to +18.3 bp, against the +8 bp required. Before costs it made +12.1 bp, and turnover cost 8.2 bp.
When the fees users pay a DeFi protocol jump, its token has further to rise, because the news is on dashboards and not yet in the price.
Why people believe it
Fees are a direct measure of how much a protocol is used, and some tokens also receive that revenue through buybacks or fee switches, which would add a mechanical buyer with a lag. Most perpetual-futures traders trade price and order flow, not protocol fundamentals.
Why it would persist The argument, as stated before testing
There are hundreds of small tokens, each with a short history and a noisy fee series, and the capacity is low, so large traders would not bother to arbitrage it.
Every day we measured, for each token with a DeFi protocol behind it, how much its fees over the last 7 days grew or shrank against the 28 days before. We then ranked the tokens into five groups, bought the fifth with the fastest fee growth and sold short the fifth with the slowest, in equal dollar amounts, for the following 24 hours. We waited one extra day before trading to allow for late publication and revisions of fee figures. Trading costs and funding were charged every day, and the first 620 of the 1,466 test days were only reported, never used to tune anything.
Validation mean net at least +8 bp per day per $1 long plus $1 short; weekly-block bootstrap 95% interval lower bound above 0; selection mean above 0; orthogonalised version above 0; mean tokens per leg at least 6.
Validation mean net below +3 bp per day, or the 95% interval includes 0.
5 bp fee plus a per-token calibrated half-spread and impact at $5,000 size, times the absolute weight change (daily turnover), plus funding paid or received at historical rates.
Signal orthogonalised to the past 7-day return (daily cross-sectional regression residual); the fee signal lagged by 30 days as a placebo (should earn about zero); daily revenue in place of fees; long leg against BTC.
The validation interval (-9.7 to +18.3 bp) includes 0 and the mean (+4.3 bp) is below the +8 bp required; the test could not detect an effect smaller than about 19 bp per day.
- Validation (846 days): net +4.3 bp a day, interval -9.7 to +18.3 bp, median +5.6 bp, worst day -1,164 bp, 11.0 tokens per leg on average. Gross was +12.1 bp and turnover cost 8.2 bp (3.9 bp left); funding was negligible (about +0.4 bp).
- Selection (620 days, October 2022 to June 2024): net +3.3 bp, interval -13.7 to +20.1 bp, 5.1 tokens per leg. Gross +8.8 bp, turnover cost 5.6 bp.
- Orthogonalised to the past 7-day return, net was +4.2 bp in validation and +2.2 bp in selection, so the result is not the past week's price move in disguise; the headline net interval (-9.7 to +18.3 bp) still includes 0.
- The 30-day-stale placebo was negative: -18.5 bp net in validation (interval -32.4 to -4.6 bp) and -14.3 bp in selection (gross -10.5 and -8.7 bp). This is not explained.
- Revenue in place of fees (a secondary variant, not decisive): +9.9 bp net in validation with an interval of -5.1 to +24.6 bp, which includes 0; +3.8 bp in selection over 232 days. This is not evidence of an effect.
Net per day is not distinguishable from zero
Mean daily net return of the long-short portfolio with its 95% weekly-block bootstrap interval. The validation interval (-9.7 to +18.3 bp) spans zero and sits mostly below the +8 bp required; the 30-day-stale placebo is significantly negative.
Chart data (bp per day)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Validation, fee growth (846 days) | 4.3 | -9.7 | 18.3 |
| Selection, fee growth (620 days) | 3.3 | -13.7 | 20.1 |
| Placebo: fee growth 30 days stale (validation) | -18.5 | -32.4 | -4.6 |
Gross edge is about the size of the turnover cost
GrossTurnover costNet
Mean daily gross return, turnover cost and net return (net also includes funding), by period and signal version. Gross is between +8.8 and +14.7 bp and turnover cost takes 5.6 to 10.2 bp of it, leaving +2.2 to +4.3 bp.
Chart data (bp per day)
| Gross | Turnover cost | Net | |
|---|---|---|---|
| Selection | 8.8 | 5.6 | 3.3 |
| Validation | 12.1 | 8.2 | 4.3 |
| Selection, orthogonalised | 9.5 | 7.1 | 2.2 |
| Validation, orthogonalised | 14.7 | 10.2 | 4.2 |
Why it failed Interpretation · not established
The measured gross edge was +12.1 bp a day in validation (+14.7 bp orthogonalised), against a daily rebalancing cost of 8.2 bp (10.2 bp), and the noise in a daily long-short return (about 200 bp) is large enough that 846 days cannot separate the resulting net edge from zero. The stale placebo's negative result and the use of revised fee history are not explained by this test; the aggregator back-fills and revises fee figures, which would flatter results, but that effect was not measured.
Lesson
A daily long-short return is noisy enough (about 200 bp) that 846 days give a minimum detectable effect near 20 bp, more than twice the +8 bp required. Computing that figure before the test, next to the pass threshold, shows what the design can and cannot see.
Scope of this result
Covers about 229 tokens with a DeFi protocol and a USD-margined perpetual, daily fee and revenue totals from a public aggregator in their current, revised form, October 2022 to October 2026 in the portfolios, a 7-day over prior-28-day growth signal, daily rebalancing and 24-hour holds. Nothing was found at this size.
A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.
1. Test design
Pre-registeredDataUniverseSignalExecutionCostsControlsSample splitPass criterionKill criterionPower2. Result
ObservationsInstrumentsNet mean95% intervalMedianWorstFinding 1Finding 2Finding 3Finding 4Finding 53. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade