Data engineering · quant research
Real-Time Market Data Filter & Alert System
Streaming pipeline that records every new Solana token launch, then mines the history for trading rules that survive out-of-sample validation.
- Year
- 2026
- Status
- Active — collecting
- Stack
- Python · asyncio · WebSockets · pandas
Pump.fun launches tokens on Solana at a rate of thousands per day, and nearly all of them go to zero. Every trading “strategy” you read about the platform is vibes and survivorship bias. I wanted an answer grounded in data I collected myself.
Collect everything, filter later
The collector holds a WebSocket connection to PumpPortal and writes every token creation event, and the bonding-curve trades that follow, into an append-only event log. Deliberately, nothing is filtered at ingest: filtering early is how you bake today's assumptions into tomorrow's dataset. The pipeline's first job is to be a faithful record.
Mining rules without fooling myself
A separate miner replays the collected history looking for entry rules that would have worked. The split is chronological — train on the first 70% of time, validate on the last 30% — because a random split leaks the future into the past in time-series data. A rule only counts if it holds on data it has never seen.
- Ingest
- PumpPortal WebSocket · asyncio
- Storage
- Append-only event log, raw and unfiltered
- Validation
- Chronological 70/30 train/test split
- Alerts
- Threshold rules on live events, human-confirmed exits
Status, honestly
The collector is built for multi-day unattended runs, and the miner's first full pass over that history will decide the first rule I trade — with a small position I can afford to lose. If nothing survives validation, that is a result too, and a cheaper one than finding out live.