{"paper_id":"2511.18578","title":"Re(Visiting) Time Series Foundation Models in Finance","authors":["Eghbal Rahimikia","Hao Ni","Weiguan Wang"],"paper_date":"2025-11-23","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2511.18578v1","page_url":"https://thequant.space/flowcharts/revisiting-time-series-foundation-models-in-finance/","scores":{"empirical_rigor":8.5,"math_complexity":6,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper conducts the first comprehensive empirical study of time series foundation models (TSFMs) for excess return forecasting in global financial markets.","quote":"This paper presents the first comprehensive empirical study of TSFMs in global financial markets.","summary":"Financial time series forecasting is central to trading, portfolio optimization, and risk management, yet it remains challenging due to noisy, non-stationary, and heterogeneous data. Recent advances in time series foundation models (TSFMs), inspired by large language models, offer a new paradigm…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The conditional law of one-step excess returns is identical across assets (homogeneous market structure).","type":"market","stated":"explicit","quote":"We further assume that the conditional law of one-step excess returns is identical across assets, that is, P(r(i),ex t+1 |r(i),ex 1:t = r) does not depend on i."},{"assumption":"The time series length T is an integer multiple of the patch length L; otherwise padding is applied.","type":"model","stated":"explicit","quote":"For simplicity, in Equation (6), we assume that T is an integer multiple of L; otherwise, the sequence can be padded to satisfy this condition."}]},"datasets":{"status":"available","items":[{"name":"Global daily excess returns","provider":null,"kind":"market-prices","access":"unknown","period":null,"universe":"94 countries","quote":"Section 5.2 broadens the scope to a global context by training all models, including both benchmarks and TSFMs, on an extensive dataset encompassing 94 countries."},{"name":"Synthetic data augmentation","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"Increasing the dataset size, incorporating synthetic data augmentation, and applying hyperparameter tuning further enhance performance."}],"known":[{"id":"jkp","name":"JKP Global Factor Data","provider":"Jensen, Kelly and Pedersen","kind":"market-prices","access":"public","quote":"For each trading date t, if all firms’ returns from a country are 10The data volumes of the U.S., global, and JKP-augmented datasets are also reported in Figure A.4, measured in gigabytes (GB), reaching over 120 GB for the JKP-augmented data in 2022, which represents the largest dataset used.","mentions":102,"used_in":10},{"id":"crsp","name":"CRSP","provider":"Center for Research in Security Prices","kind":"market-prices","access":"commercial","quote":"To maintain consistency between CRSP and Compustat records, daily returns from Compustat are winsorized using the CRSP return distribution as a benchmark.","mentions":8,"used_in":2},{"id":"compustat","name":"Compustat","provider":"S&P Global","kind":"fundamentals","access":"commercial","quote":"To maintain consistency between CRSP and Compustat records, daily returns from Compustat are winsorized using the CRSP return distribution as a benchmark.","mentions":5,"used_in":2},{"id":"kaggle","name":"Kaggle","provider":"Kaggle","kind":"other","access":"public","quote":"130 Table D.2: TSFM Pre-Training Datasets: Provenance, Composition, Volume, and Financial Data Inclusion Model Pre-training Datasets Composition (%) Number of Observations Finance-Related Datasets Chronos Publicly available datasets from domains including Monash, Kaggle, M-series, retail, weather, and finance, plus synthetic datasets gener-ated via Gaussian Process (GP) methods (KernelSynth) and Time Series Mixup (TS","mentions":1,"used_in":1}],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"daily"},"transaction_costs":{"status":"available","modeled":"yes","headline_net_of_costs":null,"detail":"Results are evaluated net of transaction costs under four scenarios: no costs (0 bps), fixed costs of 20 bps and 40 bps, and a mixed-cost structure (≈21.3 bps for small‑cap stocks and ≈11.2 bps for large‑cap stocks).","quote":"Performance is evaluated net of transaction costs under four scenarios: no costs (0 bps), fixed costs of 20 bps and 40 bps, and a mixed-cost structure where small- and large-cap stocks face different estimated trading costs","source":"front_matter","rule":"R3","candidates":{"front_matter":"yes","evidence_card":"yes","evidence_card_quoted":true},"card_detail":"Results are evaluated net of transaction costs under four scenarios: no costs (0 bps), fixed costs of 20 bps and 40 bps, and a mixed-cost structure (≈21.3 bps for small‑cap stocks and ≈11.2 bps for large‑cap stocks).","card_quote":"Performance is evaluated net of transaction costs under four scenarios: no costs (0 bps), fixed costs of 20 bps and 40 bps, and a mixed-cost structure where small- and large-cap stocks face different estimated trading costs","meaning":"costs are part of the reported headline results"},"results":{"status":"available","evaluation_type":"trading-pnl","findings":[{"finding":"Off‑the‑shelf pre‑trained TSFMs perform poorly in zero‑shot and fine‑tuning settings compared to strong benchmark models.","kind":"performance","quote":"We find that off-the-shelf pre-trained TSFMs perform poorly in zero-shot and fine-tuning settings, whereas models pre-trained from scratch on financial data achieve substantial forecasting and economic improvements."},{"finding":"TSFMs pre‑trained from scratch on financial data deliver substantial gains in both predictive and economic performance, especially with longer input windows.","kind":"performance","quote":"TSFMs pre-trained from scratch on financial data deliver substantial gains in both predictive and economic performance and are most competitive with longer input windows."},{"finding":"Fine‑tuned Chronos (large) achieves positive out‑of‑sample R² of approximately 0.46–0.48.","kind":"predictive-accuracy","quote":"Chronos (large), which, after fine-tuning, achieves positive R2 OOS (R2 OOS ≈0.46–0.48) across window"},{"finding":"TSFMs show slower performance decay over time, indicating greater temporal robustness compared to benchmarks.","kind":"statistical","quote":"TSFMs also display slower performance decay over time, suggesting greater temporal robustness."},{"finding":"Despite improvements, TSFMs remain weaker in goodness‑of‑fit; zero‑shot models yield poor R² and fine‑tuning offers only modest improvements.","kind":"negative","quote":"TSFMs remain weaker in terms of goodness-of-fit: zero-shot models yield poor R2, fine-tuning offers modest improvements, and finance-native pre-training does not fully close the gap."}],"reported_performance":null,"out_of_sample":{"method":"walk-forward","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":32,"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":"2026-11-14T16:39:00Z"},"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/"}