In a team of two, I built a forecasting pipeline for 15-minute future log returns of 100 Indonesia Stock Exchange stocks. Inputs combined 787 ticker signals into a 3,206-dimensional feature vector, including standardized returns, missingness masks, and 44 sector aggregates.
The model was a causal Temporal Convolutional Network in PyTorch with six gated residual blocks and 5.6 million parameters, using LayerNorm and GLU gating. Training used 224 chronological fit dates, AdamW, and Smooth L1 loss.
Evaluation included a zero-return baseline, a validation RMSE of 0.0057607, and dilation ablations. We also applied post-hoc global shrinkage calibration with alpha = 0.8533 to adjust predictions around the small, noisy returns in the data.