What drives DeFi token returns against ETH? Annual trajectories of 110 protocols

In this study we at cp0x ask which protocol states and DAO decisions are associated with a DeFi token beating ETH over one year. The sample is 202 observations of 110 DeFi protocols with a market capitalisation of at least 50 million USD, in three October-to-October windows from 2023 to 2026. Only 18%, 9% and 24% of tokens beat ETH, and the median token fell 27%, 60% and 42% against ETH. Fee growth measured in ETH was the only pre-registered factor that survived all pre-specified checks in the main 2023–2026 analysis, accounting for about 30% of the variation within a window. A link between the holder share of fees and returns did not replicate on windows shifted by six months, and new products, new chains and unlocks showed no significant association. The main estimates come from the 2023–2026 regime; the supplementary 2022–2023 window is small. All results are associations, not causal effects.

Key findings

• In this study (202 observations of 110 DeFi protocols, October 2023 to October 2026), only 18%, 9% and 24% of tokens beat ETH in the three annual windows, and the median token lost 27%, 60% and 42% against ETH.
• In this study, fee growth in ETH terms was the only pre-registered factor that held in all nine robustness runs, and 10% more fee growth than a peer went with about 2.5% more annual token return against ETH (coefficient 0.256, p < 0.001, n = 175).
• In this study, fee growth alone accounts for about 30% of the within-window variation in annual returns against ETH, and the top fifth of tokens beat ETH by a median 16% with fees up 24% while the bottom fifth lost 80% with fees down 68%.
• In this study, a larger holder share of fees went with better returns at equal fee growth in the main 2023–2026 windows (coefficient 0.56, p = 0.0001, n = 135), but a pre-registered replication on April-to-April windows failed (0.12, p = 0.56, n = 109).
• In this study, new product versions (coefficient 0.21, p = 0.06), new chains (0.08, p = 0.18) and unlocks of existing tokens (−0.34, p = 0.12) showed no significant association with annual token returns against ETH.
• In this study, tokens with shrinking on-chain supply looked better (issuance coefficient −1.27, p = 0.02, n = 105) and hacked protocols did about 30% worse against ETH (p = 0.02), but the first rests on a few burn tokens, the hack was only a control, and neither survives multiple-testing correction.
• In a supplementary October 2022 to October 2023 window (28 observations), the fee growth coefficient was 0.21 against 0.26 in the main windows (one-sided p = 0.058), which is weak and fragile support.

Study at a glance

Question and motivation

A holder of a DeFi token could always have held ETH instead, so every token is measured against ETH, which also removes market-wide swings.

Two earlier cp0x studies of buybacks and distributions found no robust evidence that buybacks or distributions decided by a DAO (decentralised autonomous organisation) raise prices by themselves. This study takes a one-year view and splits each token's return against ETH into fee growth, the change in the multiple the market pays per unit of fees, and supply growth.

Data

The frozen inclusion and exclusion rules are listed below.

• Candidates are DefiLlama fee entries grouped by parent protocol, with one CoinGecko id and a DeFi category. Of 2,314 groups, 717 remained.
• A group enters a window with at least 50 million USD market cap at the start and positive DefiLlama fees on at least 80 of the prior 90 days.
• Stablecoins, gas tokens of general-purpose layer 1 and layer 2 chains, a memecoin and two non-DeFi tokens were excluded before any outcome was seen (change C2), leaving 202 of 214 observations.
• Two supply artefacts (circulating supply changing more than fivefold, both in the last window, at Fulcrom and Stargate) leave the on-chain stage, so panel_stage3.csv and panel_stage4.csv hold 200 rows; three further observations with zero end-window fees are not decomposed, leaving 197.

Method

1. Price and market cap at a date are 7-day medians of CoinGecko values. Fees F are 90-day DefiLlama sums. The holder share H is holders revenue over fees, missing when no holders revenue data exist (C11).
2. The outcome y is decomposed exactly as y = g + m − s, where g is fee growth in ETH terms, m is the log change of market cap over fees (rerating) and s is the log change of circulating supply (dilution).
3. Variance shares are cov(k, y) / var(y) after demeaning by window. H1 holds if g has the largest share.
4. Regression R1 has y on g, s, log starting multiple, log market cap, ETH beta (slope of daily token returns on ETH returns over the prior 365 days), starting drawdown, category dummies and window FE, winsorised at 2.5 and 97.5 percentiles. H2 needs a negative s coefficient with p < 0.05 and H4 a negative multiple coefficient with p < 0.05.
5. R1h adds starting holder share and its change, and H3 holds if R-squared rises by less than 0.02. R2 is the same model without g.
6. Stage 2 tests new product versions (H5), new chains (H6) and treasury change (H8), positive with p < 0.05, with a hack control.
7. Stages 3 and 4 measure on-chain issuance s_issue as totalSupply minus burn address balances on Ethereum, BSC, Base, OP Mainnet and Polygon (H2a), and the unlock residual u = s − s_issue (H9). The 20 largest residuals were audited without looking at returns.
8. The replication repeats R1h on April-to-April windows from 2023 to 2026 (HR1 pooled, p < 0.05; HR2 first two windows, p < 0.10). A forward test (HF) covers 2026-10-01 to 2027-10-01.

Results

Returns against ETH

This table shows how many tokens entered each window and how they did against ETH.

The last window holds 95 of 202 observations and weighs heavily on pooled results.

Decomposition and stage 1 hypotheses

This table shows the four stage 1 hypotheses, with robustness counts over nine runs excluding the main run.

A share near 1 means returns load on fee growth, not that fees explain all of them. Fees swing about twice as much as prices and the multiple moves against them (correlation −0.89), so fee growth alone accounts for about 30% of the variation (correlation 0.55).

The fee growth coefficient of 0.256 (95% interval 0.163 to 0.349) means about a quarter of relative fee growth reached the price, which fits both market doubt about fee jumps and noise in fee data. H4 falls to p = 0.06 in a post hoc check without six observations with incomplete end-window fees.

ETH beta was not a hypothesis, yet a token with beta 1.5 did about 38% worse than one with beta 1.0 (coefficient −0.97, p = 0.001, significant in 7 of 9 runs).

Holder share and its replication

This table compares the holder share coefficient in the main windows with the pre-registered replication.

In the main windows a token passing 30% of fees to holders looked about 18% better than one passing none. Both replication tests failed, so by the rule fixed before computing, the result is found in one specification and not replicated.

The shifted windows overlap the main ones by half, so they are not an independent sample. The forward test covers 74 protocols, 67 with holder share data, and will be computed after 2027-10-08.

Stage 2. Products, chains, treasury and hacks

This table shows stage 2, with robustness over eight runs excluding the main run.

Products count only if they had fees at the window end, so failed launches are left out and even a positive sign would not show that launches help.

Stages 3 and 4. Dilution measured on-chain

A check on LDO showed CoinGecko recounting circulating supply retroactively, so these stages read supply from the chain.

The coefficient −1.27 implies about 11% weaker return for 10% annual issuance. Half the observations had no supply change, and 21 of 55 changes were reductions (NXM, Bancor, Olympus, Uniswap, Stargate). Without NXM, Bancor and Olympus the coefficient is −0.84 with p = 0.18 (post hoc), so shrinking supply looked better while an effect for inflationary tokens alone is not shown.

Lido among staking protocols

This descriptive table, outside the frozen hypotheses, shows LDO with on-chain supply next to staking peers in the sample.

Of 14 staking tokens only Sanctum beat ETH, by 1% in the last window. Lido's share of 90-day fees among the ETH liquid staking products listed in data/eth_lst_share.csv was 88%, 74%, 50% and 53% in October 2023 to 2026, and the LDO market cap to revenue ratio fell from 43 to 12. These shares differ from the Lido share in a separate cp0x study (89%, 84%, 64% and 70%), which takes the whole DefiLlama Liquid Staking category, splits it by base asset and keeps ETH only.

Robustness and replication

Supplementary window 2022–2023

This window runs from 2022-10-01 to 2023-10-01, the end of the bear market after Terra and FTX. Its method was frozen (sha256 0e0c8a55) with the entry rules of the main analysis and R1 without window FE, and counts as consistent if b_g is positive with one-sided p < 0.10.

HC3 means heteroskedasticity-consistent errors suited to small samples. The result passes the frozen rule but is fragile, as dropping LooksRare, Biswap or Osmosis alone makes it insignificant. The window is not fully blind, because median returns were seen in the survivorship check before it was chosen, though no regressions were run before the freeze. An April 2022 window was rejected because Terra ecosystem tokens that lost 60% to almost 100% against ETH drop out of the data.

Limitations

• These are associations, not causes, and fee growth is measured over the same year as returns.
• Mostly the 2023–2026 regime, with 95 of 202 observations in the last window; the supplementary 2022–2023 window adds only 28.
• About a dozen pre-registered tests with eight or nine robustness runs each, so single results may be chance.
• DefiLlama may drop dead protocols entirely. Three observations with zero end fees (Ribbon, NFPrompt, ZORA) left the decomposition after losing 78% to 85% against ETH.
• Fees depend on DefiLlama adapters. On-chain supply ignores tokens locked in treasuries or vesting and misses Arbitrum and Avalanche tokens.
• Some token mappings are debatable (Geodnet, ZORA, WLFI, Believe).

What this means for DAOs

The most reliable association with beating ETH was fee growth measured in ETH. Revenue that rises in USD only because ETH rises adds nothing for a holder comparing the token with ETH.

This study did not show which DAO decisions grow fees. Neither this study nor two earlier cp0x studies of buybacks and distributions give evidence that returning value to holders lifts a token against the market by itself, and the forward test will check the holder share again. This is not investment advice.

Frequently asked questions

How many DeFi tokens beat ETH from 2023 to 2026?

In this study, 18%, 9% and 24% of tokens beat ETH in the years to October 2024, 2025 and 2026. The median token lost 27%, 60% and 42% against ETH across 202 observations of 110 protocols.

Does fee growth drive DeFi token returns?

Fee growth in ETH terms was the only pre-registered factor that survived all pre-specified checks in the main 2023–2026 analysis, with 10% more fee growth matching about 2.5% more return. Both are measured over the same year, so this describes rather than predicts.

Does a higher holder share of fees, such as a fee switch, help?

In the main windows, tokens passing more fees to holders looked better at equal fee growth (0.56, p = 0.0001, n = 135). A pre-registered replication on shifted windows failed (0.12, p = 0.56, n = 109), and fee switch activations were not tested directly.

Does token dilution hurt returns against ETH?

CoinGecko-based dilution showed no significant association (−0.32, p = 0.14). On-chain issuance had a coefficient of −1.27 (p = 0.02, n = 105), driven mainly by burn tokens doing better, and unlocks showed no significant association (−0.34, p = 0.12).

Do the results hold in other market conditions?

A supplementary window from October 2022 to October 2023 gave a fee growth coefficient of 0.21 against 0.26 (one-sided p = 0.058, n = 28). With HC3 errors p is about 0.11, so the support is weak.

Data and code

All data, code, the frozen method and the review for this study are on GitHub.

To reproduce, follow code/README.md and re-download CoinGecko prices with fetch_prices.py.

Version history