Why 2× becomes 10×
Six cases reaching 5× or more were studied closely to understand why one doubler continues and another dies.
Precedent · Trained from scratch
The market moves.
The model learns.
A predictive market model trained from scratch on Pump.fun launches to forecast outcomes before they’re known.
Contract4xyqzM8Cyrqfghzm9j8hg1S7xr8LLgMpc4TGU4fpump
01
Independent by design
02 / Launch forecast
Before the outcome.
A clearer reading.
A model built around the launches it observed. Explore runner rankings, upside probabilities and downside risk from the first minutes.
- 01Start with an address.
Identify the contract and find any observed launch record.
- 02Read the market.
Request the latest indexed trading pair and its activity.
- 03Look beneath the surface.
Read ownership, contract permissions and liquidity records.
- 04Open the forecast.
Load recorded estimates or an illustrative forecast preview.
Market lookup
Launch forecast
Price history
Price and volume across the indexed trading window.
Retrieving price history…
Trading activity
Recent trades
Liquidity venues
Primary market
Launch record
Ownership distribution
Where the supply sits.
A closer look at the largest accounts and the share held in identified pools.
Largest token accounts
Creator position
Launch investigation
Follow the first moves.
Separate the launch record, wallet relationships and transaction evidence.
Deployment provenance
Wallet relationships
Connections returned by the network provider. Select a wallet to inspect its record.
Open this view to retrieve wallet relationships.
Transaction trail
Read the available on-chain history and decode the instructions behind the launch.
Run a trace to inspect transaction programs, transfers and any verified creation record.
Contract capabilities
Permissions and token properties returned by the underlying records.
Evidence checks
Token extensions
Reported observations
These are source observations, not a prediction of returns or a guarantee of safety.
Evidence coverage
See what was returned, where it came from and when it was retrieved.
Source records
Read the evidence in context.
Market values reflect the selected trading pair and can differ across venues. Holder concentration describes token accounts returned by the provider; it does not establish who controls them. Identified pools are excluded from the largest non-pool share.
Missing records remain unavailable. Public sources can lag a new launch, limit requests or return incomplete coverage. Retrieval times are shown in your local time.
03 · The model recordSource record · 16 Oct 2026 ↗
Built from observation.
Measured against reality.
Every launch adds context. Every prediction leaves a record. The dataset, training history and published results, brought into one view.
- 143,749
- Launches observed
- 4,983,624
- Market snapshots
- 75,426
- Predictions recorded
Homepage capture · 16 Oct 2026, 11:01 UTC152 launches being tracked · 2,014,297.4 combined tracking hours
The dataset, in detail
- Model version counter
- 70
- Experiments written up
- 21
- Deep studies
- 101
- Runners followed
- 2,084
- Snapshots with known results
- 556,782
- Tracking check-ins
- 361,079
- Predictions since the collector fix
- 74,751
- Alive launches, prior six hours
- 745
- Launch rate at capture
- 17.1 / min
- Prediction rate at capture
- 5.9 / min
- Snapshot rate at capture
- 72 / min
- Observation period
- 11 days
The homepage reports 70 model versions alongside 69 training runs. These are different source counters; the latest version, v0.70, was still training. Combined tracking hours sum time across launches.
A prediction is only the beginning.
The test is what happens next.
19%
of the highest-ranked 5% doubled within six hours.
Frozen v0.41 · Predictions locked in advance
The record, before the result.
68 of 363 top-ranked launches doubled, against 3.3% across the full cohort. The 90% interval for the top group is 15–23%.
Top 5% of model scores19%
All evaluated launches3.3%
- Doubling ranking score
- 0.801 Area under the curve
- Doubling average precision
- 0.144 Across ranking thresholds
- Doubling calibration error
- 0.011 Lower is better
- Evaluation sample
- 7,266 Launches · 8–15 Oct
Primary five-minute checkpoint, post-fix cohort. Only timely predictions on active launches with a complete six-hour outcome window are evaluated.
Explore every outcome
| Outcome | Ranking score | Average precision | Top 5% hits | All positive outcomes |
|---|---|---|---|---|
| 2× within 6h | 0.801 | 0.144 | 68 / 363 | 240 / 7,266 |
| 5× within 6h Ranked by runner score | 0.647 | 0.025 | 14 / 363 | 71 / 7,266 |
| 10× within 6h | 0.645 | 0.008 | 5 / 363 | 23 / 7,266 |
| 20× within 6h | 0.809 | 0.012 | 3 / 363 | 12 / 7,266 |
| −70% within 1h | 0.955 | 0.507 | 186 / 363 | 281 / 7,266 |
Ranking measures how well the model orders positive and negative outcomes; 0.50 is a coin flip. Average precision is a ranking measure, not a win rate. There is no separate 5× probability head.
Ranking lift does not imply trading profitability. Every paper-trading rule in the source post-fix test lost after costs.
Read the original record ↗The frozen model, coverage and calibration
One fixed model.
Version 0.41, dataset v32. Frozen on 8 Oct 2026 at 17:57 UTC; trained on launches through 7 Oct at 21:38 UTC. A fixed hash sample selected 25% of tracked launches before any outcome.
- Artifact fingerprint
- 1f982faf065c…
- Freeze fingerprint
- 68351158dbdd…
- Public repository checkpoints
- 4
The misses stay visible.
The homepage record reports 144 missed and 2,658 late predictions, alongside 75,426 recorded predictions. Twelve data bugs are documented in the original changelog.
The separately scoped forward-test primary cohort records 51 missed, 887 late, 8,466 quality flags and 6,836 incomplete windows. These are retained in the ledger and excluded from scoring.
Probability, checked.
- Forward doubling · predicted / actual
- 4.4% / 3.3%
- Forward doubling calibration error
- 0.011
- Forward collapse · predicted / actual
- 3.8% / 3.9%
- Forward collapse calibration error
- 0.012
Historical homepage collapse calibration error: 0.007. Historical and prospective results describe different evaluations.
A model that keeps learningEvery generation.
Full training history ↗Every generation.
Including the ones that failed.
- 69
- Training runs 51 succeeded · 17 failed · 1 running
- 138.83
- Processor hours Measured by the trainer
- 16
- Model questions 36 base inputs, plus feature families
v0.70
Training at capture
Another configuration. A fresh test.
Automatic search on dataset v46: 221,783 snapshots across 46,537 launches. Nine of sixteen model heads had trained when this record was captured.
- Dataset
- v46
- Started
- 15 Oct · 19:51 UTC
- Elapsed at capture
- 15h 11m 16s
- Recorded cost
- Not reported
Boosting stage 286 / 300 · training loss 0.0229 · validation loss 0.0126.
What goes into the model
Observed at a fixed age.
Launches become snapshots at 1, 5, 15, 30 and 60 minutes. Signals describe price path, volume, buy and sell flow, wallets, holder concentration, creator behaviour, socials, launch structure, fees and change since the previous snapshot.
One question. One estimate.
Sixteen specialised models cover collapse, survival, doubling, trajectory, creator exit, migration and larger runs. Boosted decision trees learn from launch observations; a separate language layer explains the outputs.
Migration benchmarking is paused in the source because the old labels were biased.
A separate test set.
The research model uses a 70 / 10 / 20 train, validation and test split by mint. Every snapshot of the same launch stays together. The v0.69 test set contains 10,012 snapshots from 4,796 launches.
Total recorded wall time: 8,450.3 minutes. Graphics-processor rental: 0 hours.
The sixteen model questions
- Collapse within 1h
- Collapse within 6h
- Survival after 1h
- Survival after 6h
- Survival after 24h
- Reaches 2× within 1h
- Reaches 2× within 6h
- Trajectory over 1h
- Trajectory over 6h
- Creator exit within 1h
- Migration within 24h Benchmark paused
- Reaches 5× within 1h
- Reaches 10× within 24h
- Upside within 6h
- Runner: 10× within 6h
- Major runner: 20× within 6h
Collapse means a low at or below 0.3× the snapshot price. Survival requires at least $10 in recent trading volume. Creator exit means selling at least half the creator holding. The frozen public ledger does not expose every research-model question.
Research-model benchmarks, including the weaker results
| v0.69 question | Reported score | Score definition |
|---|---|---|
| Collapse within 1h | 98 / 100 | Ranking area × 100 |
| Survival after 1h | 91 / 100 | Ranking area × 100 |
| 2× within 1h | 89 / 100 | Ranking area × 100 |
| Creator exit within 1h | 90 / 100 | Ranking area × 100 |
| Collapse within 6h | 98 / 100 | Ranking area × 100 |
| Survival after 6h | 93 / 100 | Ranking area × 100 |
| Direction · 1h / 6h | 60% / 61% | Balanced accuracy |
| Runner · 10× within 6h | 4% | Average precision |
| Runner · top 5% of scores | 3% | Share reaching 10× |
| Major runner · 20× within 6h | 0% rounded | Average precision |
These are the original v0.69 research-model benchmarks, distinct from both the frozen forward test and the v0.41 historical clean benchmark. Ranking scores are not accuracy percentages.
From the first commit to the frozen record
4 Oct 2026
Built from scratch
The whole system was written in one day: the collector that watches Pump.fun, the database, the trainer, the model format, the benchmarks, the ledger and this website.
The original published development history. Statements describe the source system at each recorded point in time.
Research, in the openBetter questions.
Better questions.
Stronger evidence.
What has been learned.
What needs another test.
Does a run keep going?
An early continuation signal separated runs that kept going from those that exhausted. The study contains 106 runs from one day, so the finding still needs replication.
- 0.674
- Ranking score
- 43% / 10%
- Continuation · top / bottom score groups
Next question
Does the signal hold across at least five days of post-cut-off runs?
Research findings are separate from the frozen forward test. Small samples and failed hypotheses remain visible.
Read the research ↗More questions from the research programme
Healthy sideways
A flat price with buyers, holders and volume still growing receives close tracking rather than being dropped.
Sell-offs before a collapse
A 30-minute-ahead signal in 26 cases produced a ranking score of 0.75. This remains an early research finding.
Second waves
At the homepage capture, no launch had been observed returning after a collapse in the live dataset. Tracking continued.
When to exit
Every fixed exit rule tested lost. The source describes a sellable peak lasting about a minute, with entry price setting the ceiling. The problem remains unsolved.
How the work is funded
Creator rewards received$552.31
Four recorded claims on 14 Oct 2026. The captured treasury balance was $552.31 against a $50 training target.
Recorded training spend$0.00
Compute, data and storage were each recorded at $0. Serving spend was also $0. Credit usage was not priced in dollars.
At the source capture, the operating governor reserved 135,000 data credits for forward outcomes, with 77,902 research credits remaining and 46.3 hours of forward runway.
Open the treasury record ↗These figures preserve the supplied Model Zero publication captured on 16 Oct 2026. Each evaluation keeps its original model, cohort and time boundary. They are not a live training feed.
04
New launches
The market moves.
Stay close.
Follow launches as they arrive. Choose a token to open its forecast view, then explore the market activity behind it.
The launch index
Connecting to launches…
| Watchlist | Token | Source | Market cap | Volume, 24h | Open report |
|---|
Waiting for launches.
New creations appear as the feed receives them.
Market records load from the source.
05
From observation to prediction
First the prediction. Then the proof.
Observe, learn, predict, lock and evaluate. Each stage has a place in the record.
06
Useful to know
A little more clarity.
1
What does trained from scratch mean?
The underlying research model learns its market relationships from launches it observed itself. It uses task-specific boosted decision trees, rather than another model’s pre-existing market knowledge.
2
What does the forecast measure?
Upside estimates ask whether a launch reaches 2×, 10× or 20× its snapshot price within six hours. The drawdown estimate asks whether it loses at least 70% within one hour. Runner score is a percentile rank, not a chance of success.
3
What happens when I analyze a launch?
The report opens with forecast checkpoints, then provides current market, ownership, launch and contract views. Tokens in the original archive show their recorded forecasts. Other addresses show clearly labelled illustrative forecast previews; current market observations are shown separately.
4
Are the figures live?
The model totals, benchmarks and recorded predictions come from the original Model Zero research record captured on 16 October 2026. Market prices and contract evidence are fetched when you scan. Each record identifies its source and time.
5
How is performance measured?
Held-out benchmarks and the frozen forward test remain separate. The original frozen model is v0.41, frozen on 8 October 2026. Sample sizes, missed predictions, late entries and research limitations stay visible alongside the results.
Ready to look closer?