The ML contour currently operates with 13 tokens on the USDT perpetual market, including LINK, AAVE, FET, FIL, ETH, LTC, PENGU, HYPE, DOT, BNB, DOGE, BCH, and SUI. The system focuses on intraday trading with horizons of 15, 60, and 240 minutes.
The data collection process utilizes multiple independent channels: BingX REST for closed candles and snapshots, BingX WebSocket for trades, order books, and events, Binance REST and WebSocket for price and flow, and public endpoints from other exchanges for funding rates, open interest, and positioning.
To prevent data leakage and ensure reproducibility, the system stores both exchange time and local received_ms. It uses a conservative boundary where a model is considered to have known data only after a snapshot is complete. The system also employs a a purged split method to separate training and testing data, avoiding overlapping horizons.
The scanner calculates market metrics and a fixed two-axis evaluation: D (direction from -100 to +100) and A (market activity from 0 to 100). The ML model learns the relationship between features and subsequent price movements without altering the weights of the D and A evaluations.
The initial model is based on linear regression Elastic Net, chosen for its transparency. Training begins after accumulating at least 14 days of data and 2,000 labeled rows of the current generation. The working window is limited to the last 90 days, and a candidate model must reduce the Mean Absolute Error (MAE) by at least 2% compared to a zero-return forecast.