{"paper_id":"2507.07159","title":"Large-scale portfolio optimization with variational neural annealing","authors":["Nishan Ranabhat","Behnam Javanparast","David Goerz","Estelle Inack"],"paper_date":"2025-07-09","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2507.07159v1","page_url":"https://thequant.space/flowcharts/large-scale-portfolio-optimization-with-variational-neural-a/","scores":{"empirical_rigor":8,"math_complexity":8.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes mapping large-scale portfolio optimization onto a classical Ising-like Hamiltonian and solving it with Variational Neural Annealing (VNA) using autoregressive neural networks.","quote":"We propose mapping this problem onto a classical Ising-like Hamiltonian and solving it with Variational Neural Annealing (VNA), via its classical formulation implemented using autoregressive neural networks.","summary":"Portfolio optimization is a routine asset management operation conducted in financial institutions around the world. However, under real-world constraints such as turnover limits and transaction costs, its formulation becomes a mixed-integer nonlinear program that current mixed-integer optimizers…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The covariance matrix C is positive semi‑definite, making the continuous‑valued objective convex.","type":"model","stated":"explicit","quote":"The covariance matrix C is positive semi-definite by construction, and the objective function is therefore convex."},{"assumption":"Asset returns are estimated from monthly closing prices using logarithmic returns over a five‑year look‑back window.","type":"data","stated":"explicit","quote":"Closing prices with monthly granularity are used to compute expected returns and risk terms. µ is computed as a logarithm returns, since it is time-additive, and the covariance matrix C is built with a five-year look-back window."},{"assumption":"Only assets that are present throughout the entire ten‑year trading period are included.","type":"execution","stated":"explicit","quote":"Only assets that are consistently present during the entire trading period are considered."},{"assumption":"Transaction costs consist of a variable cost proportional to trade volume and a fixed cost proportional to the number of trades.","type":"market","stated":"explicit","quote":"The transaction cost can be of two types: the first is variable cost, which is proportional to the volume of the trade, and the second is fixed cost, which is proportional to the number of trades."},{"assumption":"The VNA algorithm’s runtime scales linearly with the number of assets per gradient step (O(N)).","type":"statistical","stated":"explicit","quote":"Other factors include the computational complexity of VNA, which is of O(N) for each gradient descent step when implemented with a RNN."},{"assumption":"The normalization penalty term can be adjusted to incorporate transaction costs, assuming the penalty parameters can be made arbitrarily large.","type":"model","stated":"implied","quote":"the normalization penalty term is altered to include the transaction costs, ρn(1T x + Costv + Costf −1)2."}]},"datasets":{"status":"available","items":[{"name":"S&P 500 historical prices","provider":null,"kind":"market-prices","access":"public","period":"ten years","universe":"478 stocks (S&P 500 constituents present throughout period)","quote":"Our dataset is made up of historical prices over ten years of the S&P 500, Russell 1000 and Russell 3000 indices."},{"name":"Russell 1000 historical prices","provider":null,"kind":"market-prices","access":"public","period":"ten years","universe":"857 stocks (Russell 1000 constituents present throughout period)","quote":"Our dataset is made up of historical prices over ten years of the S&P 500, Russell 1000 and Russell 3000 indices."},{"name":"Russell 3000 historical prices","provider":null,"kind":"market-prices","access":"public","period":"ten years","universe":"2008 stocks (Russell 3000 constituents present throughout period)","quote":"Our dataset is made up of historical prices over ten years of the S&P 500, Russell 1000 and Russell 3000 indices."}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"unclear"},"transaction_costs":{"status":"available","modeled":"yes","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R1","candidates":{"front_matter":"yes","evidence_card":"unclear","evidence_card_quoted":false},"meaning":"costs are part of the reported headline results"},"results":{"status":"available","evaluation_type":"trading-pnl","findings":[{"finding":"VNA identifies near‑optimal solutions for portfolios with more than 2,000 assets and achieves performance comparable to state‑of‑the‑art mixed‑integer solvers such as Mosek.","kind":"performance","quote":"We demonstrate that VNA can identify near-optimal solutions for portfolios comprising more than 2,000 assets and yields performance comparable to that of state-of-the-art optimizers, such as Mosek, while exhibiting faster convergence on hard instances."},{"finding":"On hard instances VNA reaches a relative gap of 0.01 % in roughly half the time required by Mosek, even when Mosek’s stopping criterion is relaxed to 1 %.","kind":"performance","quote":"Increasing the relative gap requirement by two orders of magnitude still resulted in Mosek having twice the time-to-solution as VNA."},{"finding":"The average Sharpe ratios obtained by VNA are statistically indistinguishable from those obtained by Mosek across S&P 500, Russell 1000 and Russell 3000 datasets.","kind":"performance","quote":"The Sharpe ratio difference is minimal, indicating that VNA can yield comparable results to Mosek across different trading periods, sometimes with slightly better results."},{"finding":"Finite‑size scaling analysis reveals universal polynomial annealing‑time scaling with critical exponents µ = −0.58 ± 0.08 and κ = 0.16 ± 0.09.","kind":"theoretical","quote":"The critical exponents extracted from this collapse are µ = −0.58 ± 0.08 and κ = 0.16 ± 0.09."},{"finding":"VNA’s runtime per gradient step scales linearly with the number of assets (O(N)), enabling scalability to thousands of assets.","kind":"performance","quote":"Other factors include the computational complexity of VNA, which is of O(N) for each gradient descent step when implemented with a RNN."}],"reported_performance":null,"out_of_sample":{"method":"unclear","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":42,"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":"2026-11-14T17:33:27Z"},"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/"}