Mempool
Overview
A BNB Chain trading system that combines DEX activity, liquidity, wallet behavior, and sentiment into scored opportunities. Entry and exit logic changes with the detected market regime.
Problem
Early-stage tokens move too fast for manual analysis. The signals that matter — liquidity changes, wallet behavior, sentiment shifts — often happen minutes before a price move, buried in data that is abundant but untrustworthy on its own.
The system had to shorten the path from event detection to a scored, explainable decision.
Role
I designed the system architecture and the interface: the event-driven structure, the scoring model the user sees, and the translation of raw chain data into decisions a person can audit.
System
Ingestion
Mempool + DEX + external APIs
Intelligence
Scoring, clustering, sentiment
Decision
Entry / exit logic per market regime
Execution
Trade simulation and routing
Monitoring
Position tracking + PnL
Design
01
Signal over noise
The user never sees raw blockchain data. Every input is aggregated into signal strength, a confidence score, and a regime classification.
02
Decision surface, not control panel
The dashboard exists for observing signals, monitoring positions, and understanding reasoning — the user intervenes only when necessary.
03
Scores answer the real questions
Charts don't answer "is this safe?" Every token is converted into a score and a behavioral interpretation, so the answer is on screen, not inferred.
Implementation
Market noise
Problem
Raw mempool and swap data is unreliable on its own.
Approach
Multi-layer filtering: token scoring on liquidity, tax, and volume behavior; wallet clustering; cross-market confirmation. Only signals passing every filter surface.
No historical mempool data
Problem
True historical transaction flow isn't publicly available in usable form.
Approach
A simulation-based backtesting engine using real DEX price feeds, reconstructed candle histories, and liquidity-aware execution modeling with slippage and gas costs.
Overfitting risk
Problem
Optimizing parameters on past behavior produces strategies that only work in the past.
Approach
Walk-forward optimization, rolling test windows, and out-of-sample validation before any configuration goes live.
Regime instability
Problem
One strategy performs poorly across different market conditions.
Approach
A regime detector classifies the market — trending, accumulation, distribution, choppy — and maps each state to its own strategy configuration.
Result
- Noisy chain events become traceable decisions: what changed, how strong the signal is, and which strategy applies now.
- Strategies are validated on unseen data before going live, and switch automatically when the market regime shifts.
- The interface reads like a live monitoring system — continuous state, visible reasoning, intervention only when needed.