Quantum-Feature Encoding for Bidirectional LSTM Price Direction Prediction

Summary

This experiment adds features from a fixed three-qubit circuit to a bidirectional LSTM that predicts the next EURUSD hourly candle’s direction. Recent-window mean return, volatility, and range determine rotation angles; linked qubits produce an eight-outcome histogram. The method derives seven descriptors from that histogram, including entropy, dominant-state probability, the share of significant outcomes, bit agreement, and variance. The circuit runs on a simulator, so the approach uses quantum formalism as a nonlinear feature transform rather than a quantum computer or demonstrated computational advantage.

On roughly 1,500 EURUSD hourly candles, the author reports baseline accuracy of 47–52% and 62–67% after adding the features, using a small test set of about 130–160 examples. The article stresses that these results are preliminary and may shrink or disappear with another split, instrument, or timeframe. It does not establish a profitable trading strategy and notes runtime limits, non-stationarity, and the need for broader validation, ablation studies, and trading metrics.

Key ideas

• Three window statistics set the input angles for a fixed circuit whose measurement histogram yields seven model features.
• The circuit is simulated and serves as a probabilistic nonlinear transform, without a claim of quantum hardware advantage.
• The reported accuracy improved on a small EURUSD hourly test set, but the author warns the result may not generalize.
• The experiment predicts candle direction and does not demonstrate profitability after trading costs or risk.
• Suggested follow-up work includes ablation tests, alternate features, instruments, and evaluation with trading performance measures.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.