Wunder Challenge
Sequence forecasting for anonymized market features using windowed baselines and a residual BiGRU.
Built most of the data pipeline, model and CPU inference export.
128, 256 and 500-step windows
- Forecast anonymized market sequences with windowed baselines and a residual BiGRU.
- Built most of the data pipeline, model and CPU inference export.
- Year
- 2025
- Result / current state
- Trained over 128, 256 and 500-step windows with warmup normalization.
- Technologies
- Python
- PyTorch
- NumPy
- pandas
What it does
The challenge predicts the next vector of anonymized market features from 1,000-step sequences, scored by average R².
My contribution
I built a memory-mapped dataset pipeline, augmentation, warmup normalization, residual model and CPU-safe inference export.
Technical highlight
A bidirectional GRU with attention pooling predicts residuals over mean, last-value, EMA and linear-trend baselines, with hand-built window features and bf16 GPU training.
Results and status
I did not finish a final submission. The project is a substantial partial challenge solution and should be read as work in progress.