Session-Aware Cyclical Time Features for Forex Machine Learning
Summary
The document explains how to represent time in machine-learning features for forex. It argues that raw integer timestamps and hour values misrepresent cyclical adjacency, then describes encoding periodic variables with sine and cosine pairs. Adding Fourier harmonics can represent more complex intraday patterns, while increasing feature count can raise overfitting risk. The proposed feature set includes time cycles, forex session context, session-specific volatility, and calendar effects such as period boundaries.
The method accounts for the different activity patterns across Sydney, Tokyo, London, and New York sessions and their overlaps. Session times are represented with fixed UTC boundaries, an approximation that does not track local daylight-saving shifts. A timeframe-aware gate is intended to omit intraday features for coarser bars. The article outlines an implementation pipeline and cites established references, but the supplied excerpt contains no empirical performance comparison showing that these features improve forecasts or trading results.
Key ideas
- Sine and cosine pairs preserve adjacency at the boundaries of cyclical time variables.
- Higher Fourier harmonics can encode multiple peaks in intraday activity patterns.
- Session indicators and rolling session volatility can represent changing forex market conditions.
- Fixed UTC session boundaries simplify encoding but shift relative to local time during daylight-saving periods.
- Feature selection should reflect the bar timeframe to limit irrelevant inputs and overfitting.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.