Memecoin Price Forecasting on Solana and Base Chains: A Comparative Study of LSTM, CDP Agent and ELIZA OS with Interpretability Insights

Abstract

In the rapidly evolving decentralized finance landscape-where retail traders struggle to compete alongside institutional traders, this project provides a dedicated, Artificial Intelligence (AI) -powered trading assistant equipped with the ability to unlock professional(pro) trading strategies without any need to code. By increasingly integrating a trained machine learning model, where the database consists of over 1 million Solana memecoin data points, with a conversational AI interface and on-demand, blockchain-level analytics, the system enables users to trade on tokens such as Base Mainnet or Solana assets with one natural language command. The AI queries live Decentralised Exchange (DEX) data through The Graph protocol, determines matches, and generates$\mathbf{1 5}$-minute price predictions using the Long-Short Term Memory (LSTM) neural network. The AI performs all processes autonomously, allowing it to manage wallets and make trades using an independently validated Return of Interest (ROI) of 30.57 %. The platform simplifies candlestick patterns, liquidity-level data, and market indicators into chat-based task workflows over a 3-step process. The model is created with TensorFlow for predictive analytics, Collateralised Debt Position (CDP) Agents for engagement, and uses Coinbase's Software Development Kit (SDK) for trading. This research proves that AI can level the playing field for casual traders using complex algorithmic trading strategies and data through simple human engagement.