What does the market pay for in DeFi tokens? Category growth, market share, monetisation and valuation

This cp0x study asks what one-year returns of decentralised finance (DeFi) tokens against ETH were made of, what the market paid for at the start of each year, and whether tokens that looked cheap relative to their business later did better. The sample is 200 observations of protocols with a market cap of at least 50 million USD in three October-to-October periods from 2023 to 2026, replicated on 169 observations in April-to-April periods. None of the three pre-registered hypotheses held. At equal fees, protocols that kept a larger share of fees (a higher take rate) were not valued higher, with a coefficient of −0.03 and a 95% confidence interval (CI) from −0.40 to 0.34 on 120 observations, and cheap tokens did not later outperform expensive ones. These are associations, not causes. The main estimates come from 2023–2026, and the supplementary 2022–2023 window is small and pointed the other way on valuation, on about 10 informative observations.

Key findings

• In this study (October 2023 to October 2026), share growth and category growth had similar, weak associations with token returns against ETH, with coefficients of 0.20 and 0.16 and a difference of 0.04 (standard error 0.08, one-sided p 0.32, n = 168), so no difference was detected.
• In this study, the median DeFi protocol with a token lost 47%, 43% and 30% of its category fee share in the three years from October 2023 to October 2026, while median category fees grew against ETH by 49%, 72% and 7%.
• In this study, at equal fees, a higher take rate was not associated with a higher market cap at the starts of October 2023, 2024 and 2025, with a coefficient of −0.03 (95% CI −0.40 to 0.34, n = 120), which rules out valuation in proportion to revenue.
• In this study, market cap rose much less than proportionally with fees, with a coefficient of 0.37 on log fees, so a protocol with twice the fees was worth about 30% more on average.
• In this study, tokens that were cheap relative to their business did not later beat expensive ones, with a coefficient of 0.13 in October periods (95% CI −0.09 to 0.35) and 0.29 in April periods (two-sided p 0.004), the opposite of the hypothesised sign.
• In this study, Lido's share of ETH liquid staking fees was 89%, 84%, 64% and 70% in October 2023 to 2026, and in the last year its share and take rate rose while the ETH staking market and Lido's revenue multiple fell.
• In a supplementary October 2022 to October 2023 window, this study's valuation results reversed sign (take rate coefficient 0.83), but on only about 10 informative observations.

Study at a glance

Question and motivation

DAO (decentralised autonomous organisation) debates often assume a chain from fees to revenue to token value. If it holds, a protocol that keeps more of its fees or wins share from competitors should be valued higher. This study tests that chain.

Stage 1 decomposes past returns into category growth, share, other fees, take rate, rerating and dilution. Stage 2 asks what the market paid for at the start of a year, and stage 3 whether tokens cheap relative to that model later did better. It builds on an earlier cp0x study of annual token trajectories, where fee growth was the most consistent correlate of returns.

Data

Inclusion and exclusion rules.

• Observations are those of the panels from the earlier study of annual token trajectories, with the same entry rule and token exclusions.
• Categories follow the DefiLlama snapshot of 2026-10-07. Liquid Staking, Liquid Restaking and Restaking are split by base asset using the first listed chain. A group's main category is that of its product with the largest all-time fees.
• Revenue counts only if it is at least 1% of fees. Stages 2 and 3 exclude ve(3,3) exchanges (vote-escrow designs where voters receive all pool fees), Tokenlon, World Liberty Financial, Lista DAO and ApeX because their revenue is not comparable, which removes 13 main observations.
• Observations where the two DefiLlama fee sources are both positive and differ more than twofold are dropped, 3 main observations in stages 2 and 3 (start date only) and 5 in stage 1.

Method

1. Fees F, revenue R and category fees C are 90-day sums. share = F_main / C, take rate = R / F, D = F / F_main, revenue multiple M_R = MC / (R × 365/90), fee multiple M_F = MC / (F × 365/90), with MC the market cap.
2. Decompose exactly, y = Δln(C/ETH) + Δln share + Δln D + Δln M_F − Δln S, with S the circulating supply. The revenue variant adds Δln take and uses M_R. Maximum identity error is 2·10⁻¹⁵.
3. H1 regresses y on Δln(C/ETH), Δln share, log market cap, beta, drawdown and period FE, and holds if the share coefficient exceeds the category coefficient (one-sided p < 0.05, two-way clustered by protocol and category × period).
4. H2 uses model V at period start, ln MC = a ln F + b ln take + c prior share growth + category FE + period FE, and holds if b > 0 with p < 0.05, clustered by protocol. b = 1 with a = 1 means value proportional to revenue.
5. H3 regresses y on the model V residual, beta, drawdown, category FE and period FE, and holds if the coefficient is negative with p < 0.05.
6. Winsorise at 2.5% and 97.5% within period. Holm correction across the three main p-values. A hypothesis is recognised only if it also has the same sign with one-sided p < 0.10 in April periods.

Results

Verdicts

The table shows each main-period test, its Holm-adjusted p-value and the replication check.

No hypothesis held on main periods. H2 passes replication only on the one-sided reading and would fail at two-sided p 0.107.

Stage 1. What returns were made of

The table shows per-observation medians in the fee-variant sample (189 main observations), converted from log changes to percent.

Categories grew while most protocols lost share, and the data cannot tell whether it went to new entrants or to protocols without a token.

A 1% faster share growth went with about 0.20% higher return against ETH, and 1% faster category growth with about 0.16%. The test had low power and could reliably detect only a difference of about 0.2 to 0.25 or more.

Stage 2. What the market paid for

The table shows model V with all variables at period start.

At equal fees, a higher take rate was not associated with a premium, and the interval excludes b = 1. A fee coefficient of 0.37 means twice the fees went with about 30% more market cap (2^0.37 ≈ 1.29). It is not a structural price of fees, given the 50 million USD truncation, noisy fees that include payments to liquidity providers, and category FE.

Stage 3. Did cheapness predict the next year

The table regresses next-year returns on the model V residual, where positive means expensive relative to fundamentals.

H3 predicted a negative sign, and both estimates are positive. In April periods expensive tokens later did better, and without A1 the coefficient is still 0.21 (p 0.04). The residual may capture growth expectations or quality rather than market error, which was not tested.

Lido

The table shows Lido's decomposition in log points from the study panel.

In the last year Lido regained share (+10%) and raised its take rate (+27%), but the ETH staking market shrank 30% against ETH and Lido's revenue multiple fell 38%. The take rate rise matches the DAO's cut of the Curated Module node operator base fee to 3.5% (on-chain vote 195, executed 24 December 2025).

Robustness and replication

The table covers main periods unless stated. H3 shows the two-sided p because its hypothesised sign is negative.

No pre-specified check made any hypothesis significant in the predicted direction.

Supplementary window 2022–2023

A window from October 2022 to October 2023, the end of the bear market after Terra and FTX, was tested under a separate method frozen on 2026-10-07T19:35:53Z (METHOD_W0.md sha256 0e0c8a55), with 28 observations and 16 in the valuation model.

The signs are reversed. But 6 of the 9 category keys are singletons, so the model rests on about 10 informative observations in 3 categories with 4 residual degrees of freedom, and carries little weight next to the 120 main observations. The window is not fully blind, because median returns were seen in a survivorship check before the freeze, though no regressions were run in advance.

Limitations

• Stages 2 and 3 have about 100 to 120 observations, and the main estimates come from 2023–2026, while the supplementary 2022–2023 window is small and pointed the other way on valuation.
• Categories come from a 2026 snapshot, and choosing the main category from all-time fees is a mild look-ahead.
• Group fees were switched after the freeze to the same source as revenue, because about 500 products are missing from the breakdown.
• 25 observations lack revenue data, and the frozen exploratory issuance analysis was not done.
• Many protocols have a 100% take rate by construction, and multi-chain liquid staking mapping is arbitrary for Ankr and Stride.
• H2 compares protocols at one date and does not test a fee switch within one protocol.
• Many tests were run, and all results are associations, not causes.

What this means for DAOs

In 2023 to 2026 the market responded to the size of a business and its change, but did not value DeFi tokens as a simple claim on current cash flow. Keeping more of the fees did not automatically mean a higher valuation, though revenue still funds the treasury and development. Market share deserves attention, since most protocols lost share in growing categories. This is not investment advice.

Frequently asked questions

Did the market pay more for DeFi protocols with a higher take rate?

This study found no evidence of that. At equal fees the take rate coefficient was −0.03 (95% CI −0.40 to 0.34) on 120 observations in October 2023 to October 2026. The interval excludes 1, the value implied by valuing protocols on revenue alone.

How much more was a protocol with twice the fees worth?

About 30% more on average in this study, from a coefficient of 0.37 on log fees. An earlier draft said about 35% and was corrected on 8 October 2026.

Did market share matter more than category growth?

No difference was detected in this study. Coefficients were 0.20 for share growth and 0.16 for category growth, a difference of 0.04 with one-sided p 0.32, and the test could detect only differences of about 0.2 or more.

Did cheap DeFi tokens outperform expensive ones?

They did not in this study. The residual coefficient was 0.13 in October periods and 0.29 in April periods (two-sided p 0.004), so expensive tokens did better in April periods. This argues against buying tokens because they look cheap on fees and take rate.

Why did Lido's token fall when its share and take rate rose?

From October 2025 to October 2026 Lido's share of ETH liquid staking fees rose from 64% to 70% and its take rate from 4.87% to 6.21%. But in this study's data the ETH staking market shrank 30% against ETH and Lido's revenue multiple fell 38%.

Do the results hold in other market conditions?

In this study's October 2022 to October 2023 window the signs reversed, with a take rate coefficient of 0.83, but on about 10 informative observations, so it carries little weight next to 120 main observations.

Data and code

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

Paths are relative to 04-market-valuation. To reproduce, fetch the breakdown with code/fetch_raw_r5.py, check its sha256, then run code/build_r5.py and code/analyze_r5.py.

Version history