Multi-signal FX and crypto trading research
A research platform that combines price history, live order books and AI-scored news into one trading signal, with strict risk limits and honest backtests.
- Year
- 2025
- Context
- In-house R&D
- Tools
- Python, Claude API, Kraken API, OANDA, pandas
What it is
A research tool for testing whether combining different kinds of market information beats any single signal. It covers BTC, ETH and SOL, plus EUR/USD, GBP/USD and USD/JPY, on four-hour bars. It only backtests and paper trades. Live trading is switched off on purpose.
The signals
- Price history: EMA trend, risk-adjusted momentum and z-score mean reversion.
- Market activity: order-book imbalance from Kraken’s live book.
- News: headlines from RSS feeds, filtered by keyword, then scored by Claude for relevance, sentiment, confidence and time horizon for each instrument.
- Events: new entries are blocked for 30 minutes either side of high-impact releases such as CPI, payrolls and rate decisions.
Risk first
Position sizes come from the distance to an ATR-based stop. Exposure caps, a daily loss limit and a maximum-drawdown halt sit on top. A one-file kill switch closes every position and stops the bot.
Backtests you can trust
- Every decision uses only data available at the time. Fills happen on the next bar’s open, with spread, slippage and fees.
- Automated tests confirm that adding future data never changes past signals.
- Walk-forward and ablation modes show whether each signal earns its place.
- News is archived as the paper loop runs, so it can be replayed point-in-time later.
Why we built it
It’s a demanding test of how we combine data, AI and engineering discipline. The same patterns (causal features, cached LLM scoring and hard risk limits) carry straight over to client forecasting and automation work.