•
Wei, Yang & Cui (2025) embed GARCH(1,1) inside GRU and LSTM cells—joint ω, α, β plus neural gates—on SPX, DJI and NASDAQ. Not investment advice.
Takeaway. Most “GARCH + LSTM” papers are a pipeline: fit GARCH, dump σ̂ into a net. Unified GARCH-RNN puts the GARCH(1,1) update inside the cell—GRU gets an additive vol bias, LSTM a multiplicative gate—so ω, α, β stay economically readable while gates learn extra nonlinearity.
Unified GARCH-Recurrent Neural Networks in Financial Volatility Forecasting (Jingyi Wei, Steve Yang, Zhenyu Cui, 2025; arXiv:2504.09380). S&P 500, DJI, NASDAQ, 2010–2020; 2019 calm OOS and 2020 COVID stress.
Fit for learners: Don’t pipe GARCH features into a frozen LSTM. Put the GARCH(1,1) update inside the GRU/LSTM cell so ω, α, β stay readable while the net learns nonlinear persistence.
This Hub note is a plain-language intro. Not investment advice.
Plain GARCH is linear-ish and misses breaks. Bare LSTM/Transformer can forecast RV but hide persistence. Two-stage hybrids train separately, so the econometric bit cannot adapt with the net. Zhao-style embeddings exist; this paper wants vector hidden states, a GRU sibling for speed, and constraints so α+β < 1 still means something.
g_t from ω + α ε²_{t-1} + β σ²_{t-1}, project into hidden size.γ g_t into the hidden state (tanh).(1 + w tanh(g_t)).GARCH-GRU is their accuracy–speed pick in normal markets (~3× faster per epoch than GARCH-LSTM). GARCH-LSTM can edge MAE in 2020 stress. Both beat the pipeline hybrid — the point of embedding, not stacking.
It is the conceptual sequel to Hansen’s Realized GARCH and to GINN-style hybrids: structure inside the cell, not only in the loss. Read the parameter table (α ≈ 0.09–0.11, β ≈ 0.86–0.88, α+β ≈ 0.97) to see they did not throw GARCH away.
Tiny evening exercise: one step of GARCH h = w + a*e2 + b*h then h_new = tanh(h_gru + 0.1*h) with toy numbers. Learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. Index volatility bake-offs are not a recommendation to trade vol or set capital.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—gated updates are just weighted averages; then the PDF’s cell equations.