The Deep Learning Fallacy in Financial Data
Beginner algorithmic traders frequently attempt to feed raw EURUSD 15-minute candlestick closes into a 4-layer Recurrent Neural Network (LSTM). The in-sample training accuracy hits 88%, but out-of-sample forward trading performs worse than a coin toss.
Financial price series are fundamentally non-stationary: their statistical mean, variance, and autocorrelation change continuously over time. When an LSTM memorizes absolute price levels, it immediately breaks down when the market enters a new volatility regime.
What Works: Stationary Features + Gradient Boosted Trees
Institutional quantitative engineering favors algorithms like LightGBM or XGBoost trained on carefully transformed, stationary inputs:
- Log Returns & Volatility Ratios: Normalizing price movement relative to rolling standard deviation (Garman-Klass volatility).
- Order Flow Imbalance: Measuring buyer vs seller initiated volume at current bid/ask levels.
- Fractional Differentiation: Preserving long-term price memory without non-stationarity drift.