{"paper_id":"2412.09517","title":"Geometric Deep Learning for Realized Covariance Matrix Forecasting","authors":["Andrea Bucci","Michele Palma","Chao Zhang"],"paper_date":"2024-12-12","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2412.09517v1","page_url":"https://thequant.space/flowcharts/geometric-deep-learning-for-realized-covariance-matrix-forec/","scores":{"empirical_rigor":8,"math_complexity":9.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes a Riemannian‑geometry‑aware deep learning framework to forecast realized covariance matrices, including a diagonal‑block input construction and a geometric HAR extension.","quote":"we propose a novel approach for forecasting realized covariance matrices of asset returns using a Riemannian-geometry-aware deep learning framework.","summary":"Traditional methods employed in matrix volatility forecasting often overlook the inherent Riemannian manifold structure of symmetric positive definite matrices, treating them as elements of Euclidean space, which can lead to suboptimal predictive performance. Moreover, they often struggle to handle…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"Covariance matrices are symmetric positive‑definite and lie on a Riemannian manifold.","type":"model","stated":"implied","quote":"An additional fundamental limitation of all these methods is that they treat SPD matrices as if they were elements of the Euclidean space, overlooking the fact that SPD matrices inherently rely on a Riemannian manifold"},{"assumption":"The log‑Euclidean distance is an appropriate metric for the Fréchet mean and loss.","type":"model","stated":"explicit","quote":"In our approach, we first specify d(·) as a log-Euclidean distance (Arsigny et al., 2007) defined as follows"},{"assumption":"Daily realized covariance series are stationary.","type":"statistical","stated":"explicit","quote":"The covariance stationarity of each time series was validated through unit root tests, including the Augmented Dickey-Fuller (ADF), ADF-GLS, Phillips-Perron (PP), and KPSS tests."},{"assumption":"Constructing a block‑diagonal matrix of SPD lagged covariances yields another SPD matrix.","type":"model","stated":"explicit","quote":"It must be noticed that the construction of a block diagonal matrix using SPD matrices is still an SPD matrix."},{"assumption":"Portfolio evaluation assumes no short‑selling constraints unless otherwise stated.","type":"execution","stated":"explicit","quote":"GMV indicates a portfolio without no-short-selling constraints."},{"assumption":"No transaction costs are incurred in the portfolio simulations.","type":"execution","stated":"explicit","quote":"1No transaction costs are assumed."}]},"datasets":{"status":"available","items":[{"name":"Daily realized covariance matrices","provider":"Lobster","kind":"market-prices","access":"commercial","period":"June 27 2007 to July 1 2021","universe":"50 largest U.S. stocks by market cap (S&P 500 constituents)","quote":"The empirical application of our forecasting method utilizes time series data comprising intradaily close spot prices for the 50 largest U.S. stocks by market capitalization, as of November 2021, all of which are part of the S&P 500 index. This data obtained from Lobster spans the period from June 27th, 2007 to July 1s"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"unclear"},"transaction_costs":{"status":"available","modeled":"no","headline_net_of_costs":null,"detail":null,"quote":"1No transaction costs are assumed.","source":"front_matter","rule":"R3","candidates":{"front_matter":"no","evidence_card":"no","evidence_card_quoted":true},"card_detail":null,"card_quote":"1No transaction costs are assumed.","meaning":"a trading or execution result is reported without costs"},"results":{"status":"available","evaluation_type":"forecast-accuracy","findings":[{"finding":"The proposed geometric deep‑learning approach outperforms traditional methods in predictive accuracy for realized covariance matrices.","kind":"predictive-accuracy","quote":"showing that our method outperforms traditional approaches in terms of predictive accuracy."},{"finding":"The GeoHAR model with a log‑Euclidean Fréchet mean achieves the best forecasting performance among all tested models.","kind":"performance","quote":"In particular, a model with ten lags and trained using the logarithmic Euclidean distance as a loss function and the GeoHAR using a log-Euclidean-based"},{"finding":"ReSPDNet trained with log‑Euclidean loss and 10 lags (ReSPDNet‑LE10) is the best‑performing model in the statistical selection matrix.","kind":"performance","quote":"with a log-Euclidean loss and 10 lags, which is also the only model in the SSM."},{"finding":"Geometric approaches (e.g., GeoHAR‑LELE) achieve low portfolio turnover while maintaining low risk, outperforming Cholesky‑based strategies that have high turnover.","kind":"performance","quote":"some geometric approaches, such as the GeoHAR-LELE, manage to maintain relatively low turnover while delivering near‑optimal risk metrics."},{"finding":"Using a Euclidean loss for training SPD networks leads to poorer predictive results compared to geometric‑aware loss functions.","kind":"negative","quote":"Overall, our findings demonstrate that using an Euclidean loss function to train a neural network for SPD matrices leads to poor predictive results."}],"reported_performance":null,"out_of_sample":{"method":"walk-forward","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":40,"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":"2026-11-14T17:50:09Z"},"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/"}