Hull Tactical Market Prediction with Temporal Attention
Custom Temporal Attention mechanism in PyTorch for market regime prediction, with a differentiable Sharpe Ratio Loss that optimizes directly for risk-adjusted returns.
Problem
Standard MSE-based training objectives for market prediction don't align with the actual goal of maximizing risk-adjusted returns.
Approach
Implement a custom Temporal Attention mechanism from scratch to capture non-linear market regimes, paired with a differentiable Sharpe Ratio Loss backpropagated end-to-end.
Architecture
Time-series input → custom Temporal Attention encoder → prediction head. Loss: differentiable Sharpe Ratio = mean(returns) / std(returns).
Results
15% improvement over MSE baselines in backtesting.
Lessons learned
Loss function design is the most impactful choice in financial ML — optimizing for what you actually care about (Sharpe ratio) beats optimizing a surrogate (MSE).
Next
Fed-BLEND — Federated Conformal Prediction for VLMs
Novel federated conformal prediction method that mitigates hallucinations in federated fine-tuned vision-language models via abstention.
GCTAF — Time-Series Forecasting & Flare Risk Classification
Attention-based forecasting of magnetic field trajectories combined with supervised contrastive learning for solar flare classification.