All projects
AI Trading Desk
Research tool that predicts, per closed bar, the next-24-bar slope and the probability that the first take-profit is hit before the stop, for crypto, gold and forex through one pipeline. Research and paper trading only.
Python, LightGBM, ccxt, yfinance, Parquet, Bun, LLM sentiment
The problem
Trading signals rarely show their calibration or what costs do to them. This tool turns price history into a probability-weighted decision (entry, stop, three targets or no trade) and keeps the evidence needed to check whether it works.
My role
Responsibilities
Sole developer: data ingestion, feature and label engineering, model training and walk-forward validation, the news-sentiment layer and the prediction log.
Main workflows
- 1.Fetch bars from crypto exchanges (ccxt) or Yahoo Finance and store them as Parquet
- 2.Train a LightGBM model per market and read the walk-forward report: AUC, calibration and average R after costs
- 3.Pull the last 24 hours of headlines from Google News RSS and score them with an LLM
- 4.Predict on the latest closed bar; a trade is vetoed when strong news sentiment opposes it, and every prediction is logged
Architecture
- One pipeline for crypto, gold and forex, with market-specific cost settings in a single config
- Indicator features (ATR, RSI, ADX, EMA, Bollinger, higher-timeframe trend) and triple-barrier style labels for take-profit 1 to 3 before stop
- LightGBM with walk-forward validation, probability calibration and cost-aware expected value
- A Bun script scores headlines through an LLM SDK, and a separate risk manager acts as a veto layer (not wired into prediction yet)
Security and deployment
- Research and paper trading only: the system does not place orders
- The news layer is not backtested; each prediction is appended to a CSV with its sentiment value so its usefulness can be measured later
- News older than 12 hours is ignored, and results are documented as a hypothesis to paper-test, not advice