•
Petneházi (2019) turns a dilated causal CNN into a VaR machine with quantile (pinball) loss, and shows joint training on many stocks beats one-ticker nets. Plain-language intro. Not investment advice.
Takeaway. Swap MSE for pinball (quantile) loss and a 1-D dilated causal CNN becomes a VaR model: it scores the θ-quantile of tomorrow’s return, not the mean. The punchline is data pooling—one net on 100 stocks beats a net trained on a single name.
Quantile Convolutional Neural Networks for Value at Risk Forecasting (Gábor Petneházi, 2019; arXiv:1908.07978). Daily log-returns, NASDAQ/NYSE/AMEX names, 2009–2018 (repeat 1999–2008).
Fit for learners: Dilated causal CNN + pinball loss = QCNN; joint training across many stocks beats single-name nets on VaR hits.
This Hub note is a plain-language intro. Not investment advice.
GARCH + normal (or t) maps a variance forecast into a quantile by assuming a shape. Tails are rare, so one ticker has few breaches to learn from. Linear quantile regression (CAViaR-style) is honest about quantiles but thin on local temporal patterns. A CNN trained on one stock overshoots exceedance rates.
A forecasting bake-off—not a capital-rule engine.
You see loss function = target functional. Change the loss, the same conv stack predicts a tail instead of a mean. Pooling tickers is a sample-size lesson, not a Kaggle trick.
Tiny evening exercise: implement pinball: theta * max(y-q, 0) + (1-theta) * max(q-y, 0) for theta=0.05 on fake (y, q) pairs; print which side is penalised more. Learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. Petneházi is academic VaR research—not a recommendation to size positions or regulatory capital from a CNN.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—a sliding window over a list is the 1-D conv idea without the jargon.