ML memecoin AI Agent

The system is built upon 1,048,576 data points representing the top 250 Solana memecoins. Each data point is derived from 15-minute candlestick data (open, close, high, low, and volume). Data is normalized into sequences of 30 consecutive candles, representing 7.5 hours of trading history.

Training:

A Recurrent Neural Network (LSTM) is trained on these sequences to predict the next 15-minute candle close price. Training involved 45,185 sequences with an additional 5,021 sequences for validation.

1. Machine Learning Model

Purpose: The model is trained to evaluate whether a particular candlestick sequence (chart) looks promising for a price increase or decrease.

Performance: Backtesting on the validation dataset yielded:

Total Return: +30.57% ROI (without leverage). Total Trades: 5,021. Profitable Trades: 2,784. Average Return per Trade: 0.01%.

1. AI-Powered Trading Assistant

Users interact with a chat-based AI agent (built using ChatGPT), capable of performing several any on-chain trading tasks on Base Mainnet: Wallet Management: Creates and manages wallets, waits for funding, and initiates trades. Token Discovery: Filters tokens based on user-defined criteria. For example: Input: "Find cat-themed meme coins with market caps under $100M, far from ATH, rising in the last 24 hours, with over $1M in daily volume." Output: The system converts this into a query and identifies coins matching these criteria in real-time. Prediction Visualization: Provides a price prediction graph for the user-selected coins. Trade Execution: Executes trades in real-time based on user instructions, including: Recognizing insufficient funds. Asking clarifying questions if user intent is ambiguous. Suggesting actions when predictions indicate a potential downturn.

1. Real-Time Data Integration

The platform integrates real-time data using The Graph protocol: Fetches trading pool information for specified tokens. Retrieves the last 30 candlestick sequences (15-minute intervals) for analysis. Feeds this data into the LSTM model for live predictions.

1. User-Friendly Experience

Users don’t need technical expertise to trade. The natural language interface allows casual traders to: Query the market. View predictions visually. Execute trades seamlessly.

1. Key Technology

Workflow

User Query: The user inputs a natural language request, e.g., "Show me cat tokens rising in price but still far from their ATH with more than 1 mm in volume."

AI Processing: Filters coins based on user intent (RAG). Fetches trading pool data and recent candlesticks using The Graph. Feeds normalized sequences into the ML model for prediction.

Prediction Output: Displays a prediction graph for the next 15 minutes.

Trade Execution:

User selects a coin after getting the forecasts and specifies trade parameters CDP SDK. Very much possible: AI executes the trade, monitors it, and suggests actions (e.g., selling if predictions turn bearish).

Machine Learning: LSTM RNN for time-series forecasting using TensorFlow.

Coinbase Developer Platform (CDP) SDK: Wallet management, trading, and blockchain interactions as well as natural language chatting with the user.

Data Querying: Uses Retrieval Augmented Generation (RAG) to translate user input into actionable filters for token discovery.

Real-Time DEX Data: Construct the 30 15-minute candle sequence of the tokens chosen based on user intent that the model expects, I use The Graph to query every Uniswap v3 pool the tokens trade on and construct the 15-minute candles.

Trades happen on Base Mainnet and are performed by the Coinbase SDK