FX trading bot
Tested 15 years of FX data, found no reliable edge, shipped a risk-first paper bot.
- Role
- Researcher & builder
- When
- 2026
- Stack
- Python, SQLite, MetaTrader 5 bridge
Breakdown
Trading bots are sold with backtests that look perfect. Before risking real money I wanted to know whether a real, repeatable edge in currency markets exists at all.
- Backtests often hide costs and look-ahead bias
- Retail spreads and swaps eat small edges
- One bug in risk logic can wipe an account
Research first, then build only what the evidence supports. I tested classic setups, momentum, time-of-day effects, carry and cross-currency strategies on about 2.8 years of hourly and 15 years of daily data across six major pairs, then built the bot around risk controls.
- Interest-rate data lagged to avoid look-ahead
- Costs modelled before judging any signal
- Go-live rules written down before testing
Built
- Clear result: no tradable edge in the tested strategies
- Bot runs in paper mode by default with daily loss cap, drawdown halt and leverage limit
- Testing found and fixed two real bugs, including a drawdown check that could never trigger
Not yet
- No live money: best signal is a hypothesis, not a proven edge
- Go-live gate requires 26 weekly demo runs and net Sharpe of at least 0.4
How it works
Price and interest-rate data for six major pairs.
Each strategy checked with realistic costs.
Only signals that survive go forward, here none fully did.
Paper bot with kill switches waits for the go-live gate.