{"paper_id":"2310.01063","title":"Combining Deep Learning and GARCH Models for Financial Volatility and Risk Forecasting","authors":["Jakub Michańków","Łukasz Kwiatkowski","Janusz Morajda"],"paper_date":"2023-10-02","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2310.01063v1","page_url":"https://thequant.space/flowcharts/combining-deep-learning-and-garch-models-for-financial-volat/","scores":{"empirical_rigor":8,"math_complexity":7.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes a hybrid approach that combines GARCH econometric models with GRU deep learning networks for forecasting financial volatility and risk.","quote":"In this paper, we develop a hybrid approach to forecasting the volatility and risk of financial instruments by combining common econometric GARCH time series models with deep learning neural networks.","summary":"In this paper, we develop a hybrid approach to forecasting the volatility and risk of financial instruments by combining common econometric GARCH time series models with deep learning neural networks. For the latter, we employ Gated Recurrent Unit (GRU) networks, whereas four different…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The GKYZ price‑range estimator (based on opening, high, low, close) is used as the target volatility measure.","type":"model","stated":"explicit","quote":"As the main volatility estimator, also underlying the target function of our hybrid models, we use the price-range-based Garman‑Klass estimator, modified to incorporate the opening and closing prices."},{"assumption":"The estimator is scaled to match the magnitude of GARCH volatility estimates using the formula of Fiszeder (2005, 2009).","type":"model","stated":"explicit","quote":"To that end, the following formula proposed by Fiszeder (2005, 2009) is employed"},{"assumption":"Daily logarithmic returns are the input data for both GARCH and GRU components.","type":"data","stated":"explicit","quote":"Models are tested using daily logarithmic returns on the S&P 500 index as well as gold price Bitcoin prices"},{"assumption":"GRU network training uses Adam optimizer with learning rate 0.0009 and MSE loss against GKYZ forecasts.","type":"model","stated":"explicit","quote":"For the network optimisation, we use Adam optimiser (Kingma & Ba, 2017), with the learning rate set to 0.0009. The loss function is defined as the mean square error between the GKYZ volatility estimates at time t+1 and the network output"},{"assumption":"Return forecasts are generated only by the underlying GARCH model, not by the hybrid GRU component.","type":"model","stated":"explicit","quote":"GARCH model, and thereby are not further processed through the hybrid GARCH‑GRU structure."},{"assumption":"Statistical significance of MSE differences is assessed with Diebold‑Mariano test (Harvey et al. modification).","type":"statistical","stated":"explicit","quote":"Differences between the MSEs for two competing models are tested for their statistical significance via the Diebold & Mariano (1995) test, with a modification proposed by Harvey, Leybourne & Newbold (1997)."}]},"datasets":{"status":"available","items":[{"name":"S&P 500 index daily returns","provider":null,"kind":"market-prices","access":"public","period":"2009-04-06 to 2020-12-31","universe":"S&P 500 index","quote":"the S&P 500 index (5‑day week, quotations over 6 April 2009 to 31 December 2020)"},{"name":"Bitcoin (BTC/USD) daily returns","provider":null,"kind":"crypto-onchain","access":"public","period":"2013-08-05 to 2020-12-31","universe":"Bitcoin","quote":"Bitcoin (BTC/USD; 7‑day week, quotations over 5 August 2013 to 31 December 2020)"},{"name":"Gold (XAU/USD) daily returns","provider":null,"kind":"market-prices","access":"public","period":"2010-07-08 to 2020-12-31","universe":"Gold","quote":"gold (XAU/USD; 5‑day week, quotations over 8 July 2010 to 31 December 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":"Hybrid GARCH‑GRU models achieve substantially lower MSE in point volatility forecasts compared with pure GARCH models.","kind":"predictive-accuracy","quote":"combining GARCH and GRU models largely improves volatility forecasts. Enhancing a GARCH model with a GRU network results in a one‑ or two‑order‑of‑magnitude drop in MSE"},{"finding":"The improvement in volatility forecasting does not consistently translate into superior VaR and ES forecasts; the best risk‑forecasting specification varies by asset and confidence level.","kind":"performance","quote":"the gains from the volatility prediction by means of the hybrid GARCH‑GRU structures do not appear to translate unanimously into superior Value at Risk and Expected Shortfall forecasts"},{"finding":"For S&P 500 and gold, EGARCH‑GRU (skewed t for S&P 500, symmetric t for gold) yields the lowest volatility MSE; for Bitcoin, APARCH‑GRU with normal distribution is best.","kind":"predictive-accuracy","quote":"for each of the three assets under consideration, it was a hybrid model that emerged superior for the point volatility forecasting (in terms of MSE): EGARCH‑GRU models for S&500 and gold (under a skewed and a symmetric t‑distribution, respectively), while APARCH‑GRU (with a normal distribution) – for Bitcoin"},{"finding":"Risk‑assessment at 5% and 1% tolerances is best served by different hybrid specifications per asset (e.g., GJR‑GRU for S&P 500 at 5%, AP‑GRU for Bitcoin at 5%).","kind":"performance","quote":"S&P 500: GJR(SSTD)-GRU at the 5% tolerance, and AP(SSTD) at 1%; Bitcoin: AP(N) at 5%, and AP(STD) at 1%; gold: G(STD)-GRU and E(N)-GRU at 5%, and G(SSTD) at 1%"},{"finding":"Statistical tests (Diebold‑Mariano) show that many MSE reductions are significant, though some hybrid‑vs‑pure comparisons are not statistically significant.","kind":"statistical","quote":"the difference appears either statistically insignificant (conditionally normal EGARCH vs. EGARCH‑GRU) or at least not as statistically significant as one could expect (conditionally t‑distributed EGARCH vs. EGARCH‑GRU, and APARCH vs. APARCH‑GRU with a skewed t‑distribution)"}],"reported_performance":null,"out_of_sample":{"method":"walk-forward","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":45,"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":"2026-11-14T18:21:23Z"},"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/"}