Do token buybacks move prices? An event study of 43 DeFi decisions

The cp0x buyback event study measured how DeFi tokens moved against ETH after 43 dated DAO decisions to return value to holders in 16 protocols between 2022 and 2026. Each event was compared with two controls, a fixed basket of three DeFi tokens and the median of up to three tokens in a similar state before the event, and a claim was accepted only when both controls gave the same sign. Among buybacks, only the first public announcement of a programme was followed by a positive 30-day result on both controls, a median of +4.2 percentage points against the basket and +12.0 points against matched tokens, on seven events with a spread of about 40 points. Governance votes showed no positive reaction (-0.8 and -1.8 points), and late changes of known mechanisms showed a small negative (-4.3 and -3.6 points). Programme size did not order outcomes, no premium for burning was visible, and all results are associations in a small sample, not causal effects.

Key findings

• In this buyback event study (16 DeFi protocols, 2022 to 2026), the first public announcement of a buyback programme was followed by a median 30-day return against ETH of +4.2 percentage points versus a control basket (n=7, IQR 40.6) and +12.0 points versus matched tokens (n=6), the only buyback stage positive on both controls.
• In the same study, governance votes approving buybacks showed no positive reaction, with median 30-day results of -0.8 percentage points against the basket and -1.8 points against matched tokens (n=7 on each).
• Late activations, parameter changes, cuts and pauses of known buyback mechanisms were followed by a small negative 30-day median of -4.3 percentage points against the basket (IQR 3.9, the narrowest of any stage) and -3.6 points against matched tokens (n=7).
• For the first actual buyback purchases the controls disagreed (-3.8 percentage points against the basket, +17.6 against matched tokens), and a placebo test on 1,000 random date sets gave one-sided p=0.64 and p=0.04, so the study does not establish a special reaction to first purchases.
• Across five buyback launches with flows from about 5,500 to 142,000 USD per day, none of three size measures (USD per day, share of revenue, share of trading volume) ordered 30-day outcomes, and the smallest programme, ether.fi, had the best result.
• Buyback events that burned the bought tokens showed no premium, with a 30-day median of -13.3 percentage points against the basket and -5.3 points against matched tokens (n=5), and all four burn votes and activations were negative on both controls.
• The study found no evidence that a token's 90-day run-up before an event predicts its result after it, since the correlation had opposite signs on the two controls (+0.04 on 39 events, -0.21 on 35 events).

Study at a glance

Question and motivation

DAOs try to support their token in three ways. They buy the token with protocol revenue, pay revenue to holders, or cut emissions and unlock overhang. Some mechanisms are conditional and switch on only when a condition such as a revenue threshold is met.

This study asks how the market historically reacted to such decisions and at which stage a reaction appears. One decision produces several events, so the study separates the first public announcement (proposal), the governance vote (decision), the first day real money moved (flow start), and later activations, parameter changes, cuts and pauses (late changes).

Data

An event enters the computed sample only if its date is known to the day and confirmed by a governance primary source or the blockchain. Ten month-precision events and two day-precision context anchors are excluded, which leaves 43 events. Events are classed as buybacks (the "demand" class in the data), payouts or supply reductions, and coded by stage and by whether they expand or reduce value return.

Twins come from a price base of about 24 tokens, without four tokens added later for another study (JOE, BAL, YFI, VRTX). A candidate needs at least 100 observations and no event of its own from 45 days before the event to the end of the 90-day outcome window. Of 40 matchable events, 27 had three suitable twins, 9 had one or two, and 4 had none.

Method

1. Each event was dated from primary sources and coded by class, stage and direction (method/TAXONOMY.md).
2. Prices were converted into token/ETH ratios. The value on a date is the last observation on or before it, not older than 7 days.
3. The outcome is the simple return of the ratio over 30 calendar days from the event date (90 days secondary).
4. On the basket layer, the outcome is the token's return minus the median return of COMP, CRV and ENS over the same window.
5. On the twin layer, distance equals the absolute difference in drawdown from the 365-day high plus 0.5 times the absolute difference in the 90-day change. Up to three nearest candidates with distance of 0.25 or less are used, and the outcome is the token's return minus their median.
6. Results are summarised by class and stage as medians with n and the interquartile range (IQR, the 75th minus the 25th percentile).
7. Under rule 8 of the frozen method, a claim is accepted only if its sign agrees on both layers, and dropped otherwise.
8. Under rule 9, the four flow-start events with twins (Aave, dYdX, ether.fi, GMX) received 1,000 sets of random dates at least 45 days from their own events, with the same code on both layers and a one-sided p-value.
9. Slices by fate of bought tokens, execution and funding source were recomputed on both layers.
10. Rule changes after results were allowed only with a recorded reason, and the study reports none.

Results

Buyback stages on two controls

The table shows median 30-day outcomes of buyback events by decision stage on both control layers.

The proposal cell includes very strong reactions at Sky and Ethena, both with other news in the same weeks. At flow start the basket results range from -47.7 points at GMX to +46.4 points at ether.fi, and Hyperliquid, which had no twins, gained +335.6 points against the basket around its own launch.

Placebo test for first purchases

The table compares the observed flow-start median with medians from 1,000 random date sets for the same four tokens.

The layers disagree, so under rule 8 the study does not establish a special reaction to the first money spent, despite p=0.04 on the twin layer.

Payouts and supply reductions

Direct payouts and supply reductions have only one to four events per stage cell. Signs agree on both layers in every cell (for example payout proposals +2.7 points, n=4, and +9.2 points, n=2), but the study draws no conclusions from them and cannot say whether buybacks or payouts work better.

Size of the flow

The table shows five buyback launches with flow data and three size measures over 90-day windows before launch.

No measure orders the outcomes on either layer. The range is narrow, so the study says nothing about programmes as large as the Hyperliquid fund at about 6% of volume.

Run-up, direction and business state

The table shows correlations between the token's 90-day change against ETH before an event and its 30-day outcome after it.

Values use the run-up measured over 90 calendar days. An earlier version measured it over the last 90 price observations, which gave +0.10 and -0.08 for all events and +0.23 and -0.12 for decisions and activations; the correction of 2026-10-08 did not change the signs or the conclusion.

Opposite signs mean no stable link. Decisions that reduced value return were not worse than expanding ones (four reducing events at -3.9 and -1.9 points, nineteen expanding at -6.1 and -5.3 points). Revenue growth after launch did not track outcomes either, as GMX revenue rose 115% with the worst launch outcome and Jito revenue fell 53% with a positive one.

Fate of bought tokens, execution and funding source

The table shows 30-day medians of buyback events grouped by three coded variables.

Burning is the lowest fate on both layers. No premium for automatic execution and no penalty for treasury-funded buying is visible, but three treasury events, two of them Sky, make the last point a caution rather than a result.

Lido accumulation mandate case

The recomputed Lido LDO purchase mandate of 2026 gave mixed results across five closed windows, from -31.5 to +18.5 points against twins, with three signs robust to the control and two not. The windows overlap and two contain the Kelp bridge hack, so the study concludes that the data are insufficient for a conclusion in either direction.

Robustness and replication

The table shows how claims of the frozen version changed in version 2, in percentage points over 30 days.

Limitations

• Cells are small, from one to seven events, in a sample of 43 events in 16 protocols.
• The data are observational, so the study describes accompanying returns and does not establish causes.
• The basket has three tokens, the twin pool about twenty, and 4 of 40 matched events had no suitable twin.
• Many events coincided with other news, such as Maker's savings rate rise to 8% next to its June 2023 buyback announcement (+63.7 points against the basket).
• Volumes come from one venue per token, and windows start at the event-day close, so same-day jumps are missed.
• Continuous mechanisms without a discrete event are not measured, and no conditional mechanism actually launched in the sample.
• Only the version 2 method was frozen in advance, and that freeze followed an audit of earlier results.

What this means for DAOs

The study does not support presenting a buyback, payout or burn to a community as a way to raise the token price. The one stable positive signal is the reaction to the first buyback announcement, which is wide and does not repeat on later changes.

Votes, the start of purchases and larger budgets showed no reliable price reaction in the observed range, and burning showed no premium. Form and size are better argued on other grounds, such as transparency and treasury discipline. This is research, not investment advice.

Frequently asked questions

Do token buybacks raise the price of DeFi tokens?

The cp0x event study of 43 DeFi value-return events in 2022 to 2026 found no robust evidence that buybacks raise the price by themselves. The only stage positive on both controls was the first public announcement, with a 30-day median of +4.2 percentage points against a basket and +12.0 against matched tokens on seven events.

What happens to the token price after a DAO votes for a buyback?

In seven buyback votes in the study, the median 30-day return against ETH was -0.8 percentage points against the basket and -1.8 against matched tokens. The spread was about 27 to 28 points, so single outcomes varied widely.

Do bigger buyback programmes work better?

Among five launches with flows from about 5,500 to 142,000 USD per day, no size measure ordered 30-day outcomes. ether.fi, the smallest at 5,518 USD per day, had the best result on both controls, while Aave, 26 times larger, did worse on both.

Is burning bought tokens better than keeping or distributing them?

The study saw no premium for burning. Five burn events had a 30-day median of -13.3 percentage points against the basket and -5.3 against matched tokens, the lowest of three fates.

Why does the study use two controls?

Nearly equally similar twins often behaved very differently after an event. For the dYdX launch, candidates at distances 0.044 and 0.051 gave results of -75.6 and +11.3 points. Version 2 therefore uses the median of three twins next to a fixed basket and accepts a claim only when both agree in sign.

What was corrected in version 2?

The frozen version of 2026-10-02 chose twins with metrics counted in data rows rather than calendar days. The corrected method, frozen on 2026-10-04 before computation, confirmed two buyback claims and dropped the claim that first purchases gave the best reaction (+18.8 points before correction).

Data and code

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

To reproduce, download prices with code/collect_prices.py (prices are not redistributed under CoinGecko terms) and run the scripts in the order given in code/README.md.

Version history