Projects

Trading & finance

Phoenix V7

An autonomous gold-trading engine with a built-in discipline system

· 2 min read

period
2026
role
Solo build
status
Private research project

A fully automated XAUUSD (gold) trading engine that blends a machine-learning model with news, social sentiment and macro signals, then enforces risk limits it won’t let me override. Its “Guardian” layer treats the human as the biggest risk in the system.

Phoenix is the seventh generation of my automated trading research. On every 30-minute candle it pulls recent price data from MetaTrader 5, computes 14 technical indicators and asks a machine-learning ensemble (RandomForest and GradientBoosting, validated walk-forward and retrained daily) for the probability that the next move is up.

That prediction is one input, not the decision. A background thread gathers Reddit sentiment with VADER, news headlines and a macro “regime” classification, and an intelligence aggregator fuses everything into a single conviction score. Low conviction blocks the trade, conflicting signals halve the position, and economic-calendar events, wide spreads and session hours are all filtered out.

The part I am proudest of is the Guardian system, and it comes straight from security thinking: assume the operator is the threat. Circuit breakers halt trading at a 5% daily loss or a 15% drawdown from peak, a monitor closes any trade it didn’t open within about 500 ms (three attempts in a day trigger a 24-hour lockdown), and an optional kiosk mode locks the screen during market hours behind a dead-man’s switch, so emotional manual trades simply can’t happen. Every action is logged and reported to Telegram.

It is a personal research project with real risk controls, not a product or financial advice.

Highlights

  • ML ensemble with walk-forward validation and daily retraining
  • News, social sentiment and macro signals fused into one conviction score
  • Guardian layer: circuit breakers, unauthorised-trade sniper, kiosk lockout
  • Telegram alerts and automatic weekly strategy reviews

Built with

  • Python
  • scikit-learn
  • MetaTrader 5
  • VADER NLP
  • Telegram Bot API

What it taught me

Security thinking transfers. Treating the operator as a threat actor, with least privilege, fail-safes and audit logs, made the system far more robust than any extra indicator.

The code for this project is private, so I’ve described how it works rather than linking to it. Nothing here is financial advice.