•
Petneházi (2019) forecasts VaR with a dilated causal CNN + pinball/quantile loss, trained jointly across many stocks. Single-name QCNN overshoots; the joint net is the point. Not investment advice.
Takeaway. QCNN is a 1-D dilated causal CNN with quantile (pinball) loss instead of MSE—so the same WaveNet-style stack forecasts a VaR quantile, not a mean. Tails are rare on one ticker; train jointly on many stocks.
Quantile Convolutional Neural Networks for Value at Risk Forecasting (Gábor Petneházi, 2019; arXiv:1908.07978).
Fit for learners: Swap MSE for quantile (pinball) loss on a 1-D dilated CNN; pooling many names beats a per-ticker net when tails are rare.
This Hub note is a plain-language intro. Not investment advice.
VaR is a quantile of the loss distribution. Variance + Gaussian (or GARCH-normal) assumes a shape. One stock’s crash days are too few to train a hungry net. A per-ticker QCNN in this paper overshot exceedance rates—sometimes badly.
The joint net hit target exceedances with cheaper average VaR and fewer DQ rejections. Single-name QCNN did not.
A forecasting bake-off—not a desk limit.
If you can train a 1-D conv on returns, changing the loss is the smallest step from “predict the mean” to “predict a tail”. Pooling names is the data story.
Tiny evening exercise: implement pinball: theta*e if e>=0 else (theta-1)*e on toy residuals. Learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. QCNN is academic VaR research—not a recommendation to set limits or trade.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—asymmetric weighted errors are if/else arithmetic; then you can read a quantile CNN.