Systems Engineering and Quant Internships

Derive the hidden-state recurrence of a vanilla RNN and explain why long sequences can produce vanishing or exploding gradients.

Stream sequence-sized row groups from Parquet into PyTorch without loading the full competition dataset into memory.

Reproduce the competition's masked Global Weighted Pearson metric with a leakage-safe validation procedure.

Implement a stateful GRU that resets hidden state between sequences and emits two predictions at each required step.

Train the GRU on long streams using truncated backpropagation through time and gradient clipping.

Diagnose sequence-model overfitting by comparing training and validation behavior across time periods and instruments.

Export a stateful PyTorch sequence model to ONNX and verify numerical agreement with ONNX Runtime.

Profile, package, and validate a deterministic submission against the competition's one-CPU, 16-GB, 60-minute, and 20-MB limits.