{"paper_id":"2504.09380","title":"Unified GARCH-Recurrent Neural Network in Financial Volatility Forecasting","authors":["Jingyi Wei","Steve Yang","Zhenyu Cui"],"paper_date":"2025-04-13","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2504.09380v2","page_url":"https://thequant.space/flowcharts/unified-garch-recurrent-neural-network-in-financial-volatili/","scores":{"empirical_rigor":8.5,"math_complexity":7.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes two hybrid architectures, GARCH‑GRU and GARCH‑LSTM, that embed GARCH(1,1) volatility updates into recurrent neural network cells.","quote":"We propose two interpretable hybrid architectures, GARCH-GRU and GARCH-LSTM, that integrate the GARCH(1,1) volatility update into the multiplicative gating structure of GRU and LSTM cells.","summary":"In this study, we develop a unified volatility modeling framework that embeds GARCH dynamics directly within recurrent neural networks. We propose two interpretable hybrid architectures, GARCH-GRU and GARCH-LSTM, that integrate the GARCH(1,1) volatility update into the multiplicative gating…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"Return series are stationary (no unit roots).","type":"statistical","stated":"explicit","quote":"All return series reject the null hypothesis of the Augmented Dickey‑Fuller (ADF) test at the 1% significance level, confirming that the return series are stationary and do not contain unit roots."},{"assumption":"Conditional heteroskedasticity is present in the data.","type":"statistical","stated":"explicit","quote":"the Lagrange Multiplier (LM) test for ARCH effects strongly rejects the null hypothesis of homoscedasticity for all three indices, indicating the presence of conditional heteroskedasticity"},{"assumption":"A fixed rolling window of 22 trading days is sufficient for training.","type":"model","stated":"explicit","quote":"A fixed‑length window of 22 trading days (approximately one month) is used to train the model, which is then used to generate volatility forecasts for the next period."}]},"datasets":{"status":"available","items":[{"name":"S&P 500 Index daily closing prices","provider":null,"kind":"market-prices","access":"public","period":"2010-01-01 to 2020-12-31","universe":null,"quote":"We collect daily closing prices pt of three major U.S. stock indices, S&P 500, Dow Jones Industrial Average (DJI), and NASDAQ Composite, over the period from January 1, 2010, to December 31, 2020."},{"name":"Dow Jones Industrial Average daily closing prices","provider":null,"kind":"market-prices","access":"public","period":"2010-01-01 to 2020-12-31","universe":null,"quote":"We collect daily closing prices pt of three major U.S. stock indices, S&P 500, Dow Jones Industrial Average (DJI), and NASDAQ Composite, over the period from January 1, 2010, to December 31, 2020."},{"name":"NASDAQ Composite daily closing prices","provider":null,"kind":"market-prices","access":"public","period":"2010-01-01 to 2020-12-31","universe":null,"quote":"We collect daily closing prices pt of three major U.S. stock indices, S&P 500, Dow Jones Industrial Average (DJI), and NASDAQ Composite, over the period from January 1, 2010, to December 31, 2020."}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"daily"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R3","candidates":{"front_matter":"n/a","evidence_card":"n/a","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"forecast-accuracy","findings":[{"finding":"Both GARCH‑GRU and GARCH‑LSTM consistently outperform classical GARCH specifications, pipeline‑style hybrids, and neural baselines such as the Transformer across multiple metrics (MSE, MAE, SMAPE, out‑of‑sample R²).","kind":"performance","quote":"both models consistently outperform classical GARCH specifications, pipeline‑style hybrids, and neural baselines such as the Transformer across multiple metrics (MSE, MAE, SMAPE, and out‑of‑sample R2)."},{"finding":"GARCH‑GRU trains nearly three times faster than GARCH‑LSTM while maintaining comparable or superior forecasting accuracy.","kind":"performance","quote":"training nearly three times faster than GARCH‑LSTM while maintaining comparable or superior forecasting accuracy"},{"finding":"During the COVID‑19 stress period, both architectures retain superior forecasting accuracy and deliver well‑calibrated 99% VaR forecasts with lower violation ratios and competitive Pinball losses.","kind":"predictive-accuracy","quote":"In the COVID‑19 stress period, both architectures retain superior forecasting accuracy and deliver well‑calibrated 99% Value‑at‑Risk forecasts, achieving lower violation ratios and competitive Pinball losses relative to all benchmarks."},{"finding":"Average training time per epoch: GARCH‑GRU 665.78 s, GARCH‑LSTM 1,838.81 s, making GARCH‑GRU about 2.76× faster.","kind":"performance","quote":"The GARCH‑GRU model exhibits an average epoch time of 665.78 seconds, while the GARCH‑LSTM requires 1838.81 seconds. This implies that, on average, the GARCH‑GRU architecture trains"},{"finding":"Both models produce stable and economically plausible GARCH parameter estimates, preserving interpretability.","kind":"statistical","quote":"both models yield stable and economically plausible GARCH parameter estimates, confirming that the unified design retains interpretability and diagnostic transparency"}],"reported_performance":null,"out_of_sample":{"method":"holdout","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":47,"citations":{"cited_by_in_corpus":0,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":1,"extracted_at":"2026-11-14T17:40:59Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}