launchwatch-detector

Detects manufactured pumps on pump.fun — tokens pushed up by bots in a nearly perfect straight line and then emptied in a single block — from ordinary 1-minute candles. This is the detector behind launchwatch.dev (live dashboard, research note at /research.html, JSON feed at /data/orchestrated.json).

Pure Python, no third-party dependencies, causal (no look-ahead), MIT licensed.

Why

A recurring class of pump.fun tokens shows a fingerprint no organic market produces: nearly every 1-minute return is positive, returns have almost no variance, the buy/sell count ratio is unnaturally stable, and log-price fits a straight line with R² ≥ 0.98. In our observation (Sept 2026, ~900 tokens) close to 90% of them ended in a coordinated dump — dozens of wallets selling in the same block, price −99% inside one second. No sell rule reacting to price can escape it, so the only useful warning is before buying.

Install / run

pip install -e .            # or just copy the launchwatch_detector/ directory
python -m launchwatch_detector examples/CRC_manufactured.csv examples/JJQ_organic_second_wave.csv

Output (abridged):

{"manufactured": true, "detected_utc": "2026-09-04 21:30 UTC", "peak_gain_pct": 4.5,
 "collapse": {"utc": "2026-09-04 21:42 UTC", "drop_pct": -99.99, "sol_volume": 2781.3, "sells": 49},
 "minutes_to_collapse": 12.0, "file": "examples/CRC_manufactured.csv"}
{"manufactured": false, "file": "examples/JJQ_organic_second_wave.csv"}

API

from launchwatch_detector import early_signal, find_collapse, analyse

sig = early_signal(bars, window=8, min_r2=0.98)   # -> (minute_ms, close) or None, causal
col = find_collapse(bars, from_ms=sig[0])         # -> {"minute_ms", "drop_pct", "sol_volume", "sells"} or None
analyse(bars)                                     # both, plus peak gain and minutes-to-collapse

bars is a list of dicts with minute_ms, open, high, low, close, sol_volume, buys, sells (see examples/*.csv; buys/sells are trade counts in the minute). Any source of 1-minute candles with trade counts works — the reference deployment builds them from a pump.fun trade stream.

Definitions (same as the live site)

• manufactured: within the last window (8) bars — share of positive 1-min returns ≥ 0.9, coefficient of variation of returns < 0.5, standard deviation of buy/sell ratio < 0.15, R² of log-price vs time ≥ 0.98, total move > +2%.
• collapse: a 1-minute bar whose low is below 40% of the previous close (single-block dump), searched up to 6 h after detection.

Limitations

• Needs ≥ 8 minutes of the pattern; buyers in the first minutes are not warned in time.
• ~10–15% of detections fade instead of collapsing. Every detection is manufactured; not every one is dumped.
• Tuned on pump.fun bonding-curve tokens; thresholds may need adjustment for other venues.

Examples

examples/ contains three real 1-minute series derived from public on-chain data: CRC_manufactured.csv and ARROW_manufactured.csv (both collapsed by 44 wallets in one block) and JJQ_organic_second_wave.csv (an organic token, for contrast).

Source

Canonical repository: https://github.com/launchwatch/launchwatch-detector · mirror: https://gitlab.com/launchwatch/launchwatch-detector. Releases on PyPI as launchwatch-detector.

License

MIT — see LICENSE.

日本語

pump.fun で繰り返し現れる「演出型パンプ」（ボットが価格をほぼ完全な直線で吊り上げ、同一ブロックでプールを空にする型）を、 普通の 1 分足から検出する純 Python ライブラリです。launchwatch.dev の検出器そのもので、依存なし・ 因果的（後出しなし）・MIT ライセンス。使い方は上の英語の節と同じです。判定の定義（8 本の窓、R² ≥ 0.98、消滅＝安値が直前終値の 40% 未満）は サイトと同一です。検出には 8 分以上のパターンが必要で、検出のうち 1〜2 割は消滅せずに減衰します。