•
Xu, Liechty, Benthall, Skar-Gislinge & McComb (2024) train an LSTM with a loss that mixes true variance and a GARCH teacher—PINN logic for vol. Not investment advice.
Takeaway. GINN (GARCH-Informed Neural Network) is PINN logic for markets: an LSTM predicts variance, but the loss is a weighted mix of error vs realised variance and error vs a GARCH teacher—so GARCH’s stylised persistence acts like a physics regulariser.
GARCH-Informed Neural Networks for Volatility Prediction in Financial Markets (Zeda Xu, John Liechty, Sebastian Benthall, Nicholas Skar-Gislinge, Christopher McComb, 2024; arXiv:2410.00288).
Fit for learners: PINN-style loss: LSTM volatility regularised toward a GARCH forecast so the net overfits less than a pure RNN.
This Hub note is a plain-language intro. Not investment advice.
GARCH-family models are linear-ish and miss some nonlinear index features. Bare LSTMs overfit noisy squared residuals. You want both econometric structure and flexible function approximation.
A forecasting bake-off—not an options book.
This is the cleanest “two targets in one loss” story after you can write a weighted sum. It also shows metrics can prefer smoother paths—read the discussion before worshipping R².
Tiny evening exercise: loss = 0.01*((y-yhat)**2) + 0.99*((garch-yhat)**2) for toy numbers; change λ and see who wins. Learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. GINN is academic volatility research—not a recommendation to trade index variance.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—weighted sums of squares are just arithmetic; then you can read a hybrid loss.