Hypotheses
Hyperliquid Multi-Asset Regime-Analog Attention with Cost-Aware Positioning
Derived by the Research Lead from Kieran Wood et al., Trading with the Momentum Transformer: test whether attention to historically similar return regimes improves long/short positioning across BCH, ETC and FIL…
Hypothesis
Derived by the Research Lead from Kieran Wood et al., Trading with the Momentum Transformer: test whether attention to historically similar return regimes improves long/short positioning across BCH, ETC and FIL perpetuals using daily bars. This operationalization uses the normalized kernel-attention formulation in Section II rather than reproducing the trained attention-LSTM architecture. A Gaussian similarity kernel retrieves historical regimes and their subsequently realized returns; volatility scaling and a cost hurdle translate the forecast into positions. The supplied full text was read, including the methods and appendices. No neural-model performance claims are adopted. Research Lead choices are the universe, Gaussian kernel, five-day forecast horizon, historical memory, entry hurdle and protective exits. All signal choices are registered below; risk and execution settings remain fixed. Implement entirely with NumPy and permitted strategy helpers. Orders execute at completed signal-bar closes, with intrabar protective stops matched by the platform.
Analysis
Trading with the Momentum Transformer: An Intelligent and Interpretable Architecture
Backtest and paper results are hypothetical. Trading involves risk of loss.