💰 quantitative finance

Pump.fun Graduation Regime Windows: Survival Analysis of 832,941 Token Launches and the Social-Presence Effect

This study analyzes 832,941 Solana pump.fun token launches to reveal a sharp 3.18x decline in graduation rates between late 2025 and mid-2026, attributing the drop primarily to a shift toward zero-self-buy tokens while demonstrating that social-channel presence (particularly Telegram) significantly boosts success odds.

Original authors: Arati Uday Kamat

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Arati Uday Kamat

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine pump.fun as a massive, chaotic talent show happening on the Solana blockchain. Every day, thousands of new "performers" (crypto tokens) take the stage. Their goal is simple: get enough people to buy their ticket so they can graduate from the small, risky "bonding curve" stage and move to the big, famous "PumpSwap" stadium.

Most performers fail. They run out of time or money and never make it to the stadium. This paper is a massive study of 832,941 of these talent shows that happened between May and June 2026. The researchers wanted to answer two big questions:

What are the actual odds of making it?
What specific things make a token more likely to succeed?

Here is the breakdown of their findings, using simple analogies.

1. The Odds Have Dropped (The "Graduation Rate")

In late 2025, about 0.63% of tokens managed to graduate. Think of this as 6 out of every 1,000 performers getting a standing ovation and moving to the big stadium.

By mid-2026, that number dropped significantly to 0.198%. That's roughly 2 out of every 1,000.

The Analogy: It's like the talent show got much harder. The "success rate" fell by more than 3 times compared to the previous year.

2. Why Did the Odds Drop? (The "Self-Buy" Mystery)

The researchers dug into why the success rate fell. They found it wasn't because the "good" tokens suddenly became worse at performing. Instead, the mix of performers changed.

The "Self-Buy" Tokens: Some creators buy their own tokens right at the start to show confidence (like a singer buying a front-row ticket for themselves). These tokens have a much higher success rate (about 0.63%).
The "Zero-Self-Buy" Tokens: Most new tokens in 2026 were created by people who didn't buy their own tokens first. These tokens have a very low success rate (about 0.05%).

The Conclusion: The overall success rate dropped because the stage became flooded with "Zero-Self-Buy" tokens. If you only look at the tokens where the creator bought their own ticket, the success rate is actually the same as it was in 2025. The "bad luck" is just that there are way more of the "risky" tokens now.

3. The "Social Media" Superpower

The study looked at whether having a Twitter (X) account, a website, or a Telegram group helped. The results were shocking:

Telegram is King: Tokens that advertised a Telegram channel were almost 9 times more likely to graduate than those without one.
The "Full Stack" Effect: Tokens that advertised all three (Twitter, Website, and Telegram) had a success rate 17 times higher than tokens with none of them.
The Analogy: Imagine a talent show where performers who have a fan club waiting outside the door (Telegram) are almost guaranteed to get a standing ovation. Those who just show up with a guitar and no fan base usually get booed off stage.

4. The "First 5 Minutes" Rule

The researchers tracked exactly how long it took for a token to graduate.

The Finding: If a token is going to graduate, it happens fast.
The Analogy: It's like a sprint. 90% of the tokens that succeed do so within the first 2 minutes of the race. If a token hasn't crossed the finish line by the 5-minute mark, it's almost certainly going to lose. The "survival" of a token is decided almost instantly.

5. What Did They Actually Release?

This paper isn't just about numbers; it's about opening the door for others to check the math.

The Dataset: They released a massive list (860,000+ entries) of every token they watched, labeled with whether it succeeded or failed. They call this RED-PUMP-2026-v1.
The Framework: They created a new way to measure these trends over time, calling it the Graduation Regime Windows (GRW). Think of this like naming different "seasons" of the show (e.g., "The Marino Season" of 2025 vs. "The Red Pump Season" of 2026) so everyone can compare apples to apples.

Summary

The paper tells us that the "pump.fun" talent show is still incredibly risky (less than 1 in 500 tokens succeed). However, the odds are heavily stacked in favor of tokens that:

Have a creator who buys their own ticket (Self-Buy).
Have a strong social media presence, especially a Telegram group.
Get moving incredibly fast (within the first few minutes).

The drop in overall success rates isn't because the "winners" got worse; it's because the stage is now crowded with a much larger number of tokens that have none of these advantages.

Technical Summary: Pump.fun Graduation Regime Windows

Problem Statement
The Solana-based platform pump.fun facilitates the launch of memecoins, where a token "graduates" upon its bonding-curve reserve reaching a threshold (approximately 85 SOL), migrating to the PumpSwap DEX. Graduation is a rare event. While industry dashboards and prior academic work (e.g., Marino et al., 2026) have provided aggregate graduation rates, the literature lacks: (1) a public, record-level benchmark dataset for methodological replication; (2) the application of survival-analytic methods (Kaplan-Meier, Cox proportional hazards) to model graduation dynamics; and (3) a quantitative assessment of off-chain social-presence features (Twitter, website, Telegram) as graduation predictors. Furthermore, graduation rates have shifted significantly over time (from <2% in Q4 2024 to 0.63% in late 2025), necessitating a framework for comparing these distinct temporal "regimes."

Methodology
The authors present a survival analysis of 832,941 unique token launches observed continuously between May 8, 2026, and June 10, 2026 (the "RED-PUMP-2026-v1" regime).

Data Collection: A public-blockchain observer subscribed to the pump.fun program-account update stream. For each launch, the system recorded the mint address, timestamps (seenAt, created_timestamp), initial market capitalization (in SOL), and three Boolean indicators for the presence of a Twitter, website, or Telegram URL in the launch metadata.
Event Definition: The study tracks tokens for 24 hours. The terminal outcomes are GRADUATED (bonding curve crossed) or TIMEOUT (24 hours elapsed without graduation). TIMEOUT is treated as a right-censored event.
Statistical Framework:
- Kaplan-Meier Estimator: Used to estimate the cumulative graduation probability $F(t) = 1 - S(t)$ and generate survival curves.
- Stratification: Analysis was stratified by social-channel indicators (individual and combined counts) and initial market capitalization quartiles.
- Cox Proportional Hazards Model: A multivariate semi-parametric model was fitted to estimate hazard ratios ( $\lambda_i(t|x_i) = \lambda_0(t) \exp(\beta^\top x_i)$ ). Covariates included social indicators, log-transformed initial market cap, and log-transformed description length. An L2 penalty ( $10^{-3}$ ) was applied to stabilize estimates.
- Regime Comparison: The authors formalize the "Graduation Regime Windows" (GRW) framework to compare the current 34-day window against prior periods (Q4 2024 and the Marino et al., 2026 window).

Key Results

Overall Graduation Rate: The pooled graduation rate for the RED-PUMP-2026-v1 window is 0.198% (Wilson 95% CI [0.189%, 0.208%]). This represents a 3.18× decline from the 0.63% rate reported by Marino et al. (2026) for September–October 2025.
Social-Presence Effects: Off-chain social indicators are highly predictive.
- Launches with a Telegram channel graduate at 1.485% versus 0.166% without (an 8.94× lift). The multivariate Cox hazard ratio for Telegram is 5.40 (95% CI [4.73, 6.17], $p=6.6 \times 10^{-138}$ ).
- Launches advertising all three social channels graduate at 1.919% versus 0.110% for those with none (a 17.4× lift).
Market Capitalization and Self-Buying:
- The initial market cap distribution shows a large mass at the platform default of 30 SOL.
- The top quartile (initial mcap > 31 SOL, interpreted as creator self-buying) graduates at 0.634%, which nearly exactly matches the 0.63% pooled rate from the Marino et al. (2026) study.
- The lower three quartiles (dominated by zero-self-buy launches) graduate at only 0.053%.
Model Performance: The Cox model achieves a Harrell's concordance index of 0.858.
Temporal Dynamics: The first four days of the observation window showed artificially low rates (0.015%–0.043%) due to observer infrastructure warm-up latency. Excluding this period, the steady-state rate is 0.207%.

Significance and Claims
The paper claims four primary contributions:

RED-PUMP-2026-v1: The first publicly released, record-level benchmark dataset for pump.fun (860,213 launches) with a frozen schema, released under CC-BY-4.0 on Zenodo.
Methodological Firsts: The first published application of Kaplan-Meier survival curves and Cox proportional-hazards modeling to pump.fun graduation events.
Longitudinal Decomposition: An empirical decomposition suggesting that the cross-regime decline in graduation rates is substantially attributable to a compositional shift toward zero-self-buy launches (which have low success rates), rather than a decline in the success rate of self-buy launches (which remain stable at ~0.63%).
Standardization: The introduction of the Graduation Regime Windows (GRW) framework and the Pump.fun Graduation Index (PGI) as standardized metrics for future comparative work.

The authors explicitly state they do not claim causality regarding social channels, offering effort proxy, discoverability, and selection as non-exclusive interpretations. They also note that the 24-hour observation cap and detection lags are limitations, and the reported graduation rate elevations are predictive metrics for filter design rather than actionable trading signals without further validation.