•
Liu, Wang, Tran & Kohn (2023) put an LSTM inside realized GARCH so latent variance depends nonlinearly on lagged RV, returns, and past variance. Bayesian SMC. Not investment advice.
Takeaway. DeepRGARCH keeps Hansen’s return + realized-variance measurement pair, but replaces a linear intercept with an LSTM state fed by lagged RV, returns, and σ²—so high-frequency information can enter nonlinearly and with memory.
Deep Learning Enhanced Realized GARCH (Chen Liu, Chao Wang, Minh-Ngoc Tran, Robert Kohn, 2023; arXiv:2302.08002).
Fit for learners: LSTM intercept inside RealGARCH: keep the measurement equation, add long-memory nonlinearity for RV-informed volatility.
This Hub note is a plain-language intro. Not investment advice.
Daily GARCH sees only squared close-to-close returns. RealGARCH adds 5-minute RV but still treats latent variance as a short linear AR. Volatility clusters, leverages, and long memory sit awkwardly in that line. Pure LSTM vol models often dump the econometric skeleton (leverage function, measurement noise).
A volatility lab—not a vol-selling recipe.
You meet two equations, two noises—the same hygiene as state-space models—before stacking Keras. Realized Library RV is a concrete CSV, not a mystical factor.
Tiny evening exercise: from a list of 5-minute log-returns, sum(r*r for r in day) as toy RV; print why overnight is missing. Learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. DeepRGARCH is academic volatility research—not a recommendation to trade options or VaR limits from the model.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—accumulators and loops first; realized variance is a running sum of squares.