The Research Journey
Thousands of experiments across dozens of categories. The vast majority failed.
Most trading ideas look promising until you simulate them at 5-second resolution with realistic costs. The survivors are the ones that still make money after every pip of spread, financing, and slippage is deducted.
What We Tried
| Category | Outcome |
|---|---|
| Mean-reversion variants | Survived |
| Pair expansion | Survived |
| Breakout / momentum | Survived |
| Execution improvement | Survived |
| ML / deep learning | Dead |
| Microstructure | Dead |
| Risk management | Survived |
| Candle pattern TA | Dead |
| Intraday session trading | Dead |
Key Lessons
- Transaction costs kill most strategies. 1.5 pips per flip eliminates the majority of ideas.
- Diversification works. Adding uncorrelated strategies cut drawdown by 69%.
- The strategies that survived trade infrequently. A lot of edge is lost to transaction costs.
- Full-sample results can lie. Walk-forward testing exposed multiple false positives.
- The edge comes from recurring structural patterns in the FX market.
- Candle pattern TA is dead on contact with data. Zero predictive power after transaction costs.
Dead Ends
Categories of ideas that were tested extensively and failed.
- ML direction prediction
- Intraday session trading
- Cross-pair lead-lag
- Position scaling
- Candle patterns
- Tail risk hedges