We Analyzed 2,194 Token Launches—Here’s What Actually Predicts Success

Moltalyzer Team · March 2026 · 5 min read

Editor’s note: token signals, prediction-market, and GitHub feeds referenced here are no longer served by Moltalyzer as of 2026-07-10 — Polymarket moved to OrcaTrace (orcatrace.dev), GitHub to gitBeacon (gitbeacon.dev), and Master Intelligence to Signalis (signalis.dev); the token signal feed is offline. Moltalyzer now serves Moltbook community intelligence. This post is kept for its research findings.

We built an automated system that continuously scans token markets across Ethereum, Base, and BSC. It has been running for 36 days, analyzing every new token launch against a multi-signal scoring model and automatically backtesting the results.

The dataset: 2,194 completed backtests across 1,372 unique tokens. Here is what we found.

36
Days running
1,372
Tokens analyzed
2,194
Backtests completed
355
Rug pulls detected

What we scan

The system polls four data sources continuously, 24/7. Each source feeds into a multi-tier LLM pipeline that deduplicates, filters, and analyzes content before storing structured results:

Token markets
DexScreener across ETH, Base, BSC
1,372 tokens
Prediction markets
Polymarket signal detection
1,161 signals
GitHub repos
New repos with 5+ stars daily
1,279 repos
Tech news
10,138 articles into 864 digests
864 digests

The scoring system: 70% rules, 30% LLM

Every token is scored on a 100-point scale using a hybrid approach. The deterministic rules component (70% weight) evaluates seven categories:

CategoryMax points
Liquidity18
Transaction patterns17
Social presence15
Metadata quality15
Volume13
Token age12
Price action10

The remaining 30% comes from LLM analysis (Llama 70B), which evaluates token descriptions, risk flags, and generates free-text reasoning. The LLM weight adjusts dynamically based on its own confidence: low-confidence assessments drop to 15% weight, while high-confidence ones get up to 40%.

Backtest results

We define a “win” as a 1-hour price increase after the token was flagged. These are the results across the full dataset:

Unfiltered baseline18.6%n=2,194
Quality-filtered (score >= 45, $50K+ liq)28.3%n=371
High-confidence (score >= 60)37.7%n=106
18.6%
Baseline win rate
28.3%
Quality-filtered
37.7%
High-confidence

The jump from 18.6% to 37.7% is entirely from filtering. We are not picking winners. We are removing the garbage that constitutes the majority of new token launches.

Key finding: $50K liquidity is the strongest predictor

Across every feature we tested, minimum liquidity of $50,000 was the single strongest predictor of token viability. Tokens meeting this threshold showed a 50.9% win rateversus the 31.2% baseline—a larger lift than any other individual signal.

This makes intuitive sense. Liquidity is expensive to fake and correlates with legitimate project commitment. It is also the easiest filter to apply: one number, no LLM needed.

50.9%
Win rate with $50K+ liquidity
vs 31.2% baseline—the strongest single predictor in our dataset

Why we dropped Solana

We started scanning Solana alongside Ethereum, Base, and BSC. The data made the decision for us:

Ethereum57.1%n=46
Base51.7%n=339
BSC~35%n=83
Solana (dropped)21.7%

Solana token launches had a 21.7% win rate vs 51.7% on Base and 57.1% on Ethereum. The low barrier to launch on Solana means an overwhelming majority of tokens are noise. Rather than dilute our signal quality, we cut it entirely.

355 rug pulls detected

The system flagged 355 rug pulls out of 1,372 tokens analyzed—roughly 26% of all launches. Detection combines metadata quality scoring (contract verification, website presence, social accounts) with transaction pattern analysis (sudden liquidity removals, concentrated holder wallets).

Every flagged rug pull is excluded from the signal feed before it reaches consumers. This is not just about finding winners; it is about filtering out the scams.

Technical stack

The system is a TypeScript monorepo with 8 services, managed by systemd:

  • 3-tier LLM pipeline via OpenRouter: Llama 70B for fast summaries, Claude Sonnet 3.5 for analysis, GPT-4o for complex tasks
  • PostgreSQL + Prisma with prompt pre-aggregation (78% token reduction on digest prompts)
  • x402 micropayments on Base for API access (from $0.02/request, no API key signup needed)
  • Automated agent team: QC agent (30min cycles), engineer agent (hourly), product analyst (4hr metrics reviews)

Cost: $3–5/day on one machine

The entire system—scanning, analysis, backtesting, digest generation, and agent monitoring—runs on a single machine. Total LLM cost across all services is roughly $3–5 per day. The prompt pre-aggregation step (compressing source data before sending to the LLM) reduced token usage by 78%, which is what makes this cost feasible.

$3-5
Daily LLM cost
78%
Token reduction via pre-aggregation

Try the API

Token intelligence, prediction market signals, GitHub trending, and more. Structured JSON, per-request pricing, no signup required.

Explore Moltalyzer