{"paper_id":"2311.14735","title":"Generative Machine Learning for Multivariate Equity Returns","authors":["Ruslan Tepelyan","Achintya Gopal"],"paper_date":"2023-11-21","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2311.14735v1","page_url":"https://thequant.space/flowcharts/generative-machine-learning-for-multivariate-equity-returns/","scores":{"empirical_rigor":6,"math_complexity":8.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes a two‑part generative modeling framework for S&P 500 returns using a conditional importance‑weighted autoencoder for factors and a conditional normalizing flow for individual stock returns.","quote":"Our approach entails building two models that condition on the past: a conditional importance-weighted autoencoder (CIWAE, [3, 12]) for the factors (e.g., broad market indices, sector-level indices) and a conditional normalizing flow [28] for individual stock returns.","summary":"The use of machine learning to generate synthetic data has grown in popularity with the proliferation of text-to-image models and especially large language models. The core methodology these models use is to learn the distribution of the underlying data, similar to the classical methods common in…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"Stock returns are conditionally independent given the stock‑level features and the future factors.","type":"model","stated":"explicit","quote":"stock returns are conditionally independent given the stock-level features and the future factors."},{"assumption":"The universe consists of point‑in‑time constituents of the S&P 500, which can change over time.","type":"data","stated":"explicit","quote":"In our experiments, we use the point-in-time constituents of the S&P 500 as our universe"},{"assumption":"All operations in the portfolio optimization are differentiable, allowing gradient‑descent with Adam.","type":"execution","stated":"explicit","quote":"Since all the operations in the optimization task are differentiable, we used gradient descent with Adam to find the optimal long-short portfolio"},{"assumption":"A flat borrow cost of 0.1 bps per day is applied to each stock when modeling transaction costs.","type":"execution","stated":"explicit","quote":"we set the borrow cost to 0.1 bps, daily"},{"assumption":"Training, validation, and test periods are non‑overlapping and hyper‑parameter tuning is performed only on the validation set.","type":"data","stated":"explicit","quote":"Importantly, all of our hyperparameter tuning was performed on the validation set without referring to the test set."}]},"datasets":{"status":"available","items":[{"name":"S&P 500 point‑in‑time constituents","provider":"Bloomberg","kind":"market-prices","access":"commercial","period":"1996‑2022","universe":"S&P 500 constituents","quote":"We split our data into three pieces: the training set which contains data from the beginning of 1996 to the end of 2013, the validation set which contains data from the beginning of 2014 to the end of 2018, and the test set which contains data from the beginning of 2019 to the end of 2022.","known_id":"bloomberg"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"daily"},"transaction_costs":{"status":"available","modeled":"partial","headline_net_of_costs":null,"detail":"A flat borrow cost of 0.1 bps per day is applied in the portfolio optimization experiments.","quote":"we set the borrow cost to 0.1 bps, daily; we use this value since prior work (e.g., [6]) used 17 bps per year for borrow cost which, de‑annualized is approximately 0.06 bps.","source":"front_matter","rule":"R3","candidates":{"front_matter":"partial","evidence_card":"partial","evidence_card_quoted":true},"card_detail":"A flat borrow cost of 0.1 bps per day is applied in the portfolio optimization experiments.","card_quote":"we set the borrow cost to 0.1 bps, daily; we use this value since prior work (e.g., [6]) used 17 bps per year for borrow cost which, de‑annualized is approximately 0.06 bps.","meaning":"costs appear only as a robustness check, a flat fee without spread or impact, or outside the headline numbers"},"results":{"status":"available","evaluation_type":"trading-pnl","findings":[{"finding":"The proposed generative model outperforms GARCH and a classical factor model in joint negative log‑likelihood.","kind":"performance","quote":"We see that our model is able to outperform the GARCH models and the classical factor model by a signficiant margin."},{"finding":"Including industry and company financial features improves portfolio Sharpe ratio despite a slight NLL increase.","kind":"performance","quote":"the model including industry and company financials performs the best in terms of Sharpe ratio even though its NLL was slightly"},{"finding":"The final model achieves a Sharpe ratio of 2.49 on a long‑short portfolio, far above the S&P 500’s 0.64.","kind":"performance","quote":"With-Feature Model: Sharpe Ratio 2.49"},{"finding":"Using a conditional normalizing flow significantly improves model performance compared to a simple normal or Normal Inverse Gaussian.","kind":"statistical","quote":"using a normalizing flow was an important design decision, as replacing it with a simple normal or Normal Inverse Gaussian"},{"finding":"The model’s portfolio optimization outperforms both the classical factor model and the S&P 500 benchmark.","kind":"performance","quote":"our final model is able to outperform the classical factors model and the S&P 500."}],"reported_performance":null,"out_of_sample":{"method":"holdout","headline_out_of_sample":null}},"independent_evidence":{"status":"available","supports":[],"contradicts":[],"qualifies":[],"compared_with":25,"citations":{"cited_by_in_corpus":6,"by_independent_authors":5,"how_cited":{"used_as_baseline":0,"uses_method_or_data":1,"extends":0,"discusses_findings":0,"mentions":1,"unclassified":4},"independent_papers":[{"paper":{"paper_id":"2412.06417","title":"Systematic comparison of deep generative models applied to multivariate financial time series","paper_date":"2024-12-09","authors":["Howard Caulfield","James P. Gleeson"],"empirical_rigor":6,"page_url":"https://thequant.space/flowcharts/systematic-comparison-of-deep-generative-models-applied-to-m/"},"how":"uses_method_or_data","quote":null}],"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":"2026-11-14T18:19:17Z"},"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/"}