Make money doing the work you believe in
0.06 Sharpe vs. 1.99 Sharpe: same macro data, opposite outcomes. The difference is how you feed it to the model.
A new paper introduces HANET, a hybrid LSTM + attention architecture that conditions daily trading decisions on decades of macroeconomic context, and shows that the obvious way of using macro data is actively harmful.
Bolt monthly macro factors onto a daily LSTM the naive way → 0.06 Sharpe (worse than ignoring macro).
Feed the same factors through hierarchical cross-attention → 1.99 Sharpe, 22.5% returns on time-series momentum.
Shuffle the macro history's time order → collapses to 0.81. Temporal structure is the whole game.
Bonus: it's interpretable. During COVID-2020 the model leaned on 2008-crisis memory; during the 2022 inflation spike it reached back to the early-1980s Volcker era.
Tested across 55 futures (commodities, bonds, FX, equities) and the first deep-learning formulation of the carry task. Survives transaction costs too.
