{"paper":{"title":"Do Stationarity Transformations Actually Improve Time Series Forecasts? A Controlled Experimental Evaluation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Stationarity transformations improve time series forecasts only 18 percent of the time even when matched to the data.","cross_cats":["math.ST","stat.TH"],"primary_cat":"stat.ME","authors_text":"Bhanu Suraj Malla, Yuqing Hu","submitted_at":"2026-05-17T23:02:53Z","abstract_excerpt":"Stationarity transformations are standard preprocessing in time series forecasting, yet their actual impact on accuracy across different non-stationarity types and model families has received little controlled evaluation. We construct synthetic datasets with known properties - trend, seasonality, heteroscedasticity, and combinations - and apply fourteen transformation configurations across seven models and three forecast horizons (3,528 experiments). Stationarity is quantified via consensus ratios from ten statistical tests, and each transform-dataset pair is classified as matched or mismatche"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"For matched pairs, transforms improve forecasts only 18% of the time. The primary exception is variance stabilization: log and Box-Cox on heteroscedastic data improve accuracy in 60-65% of cases. Mediation analysis confirms that while transforms achieve trend stationarity, this does not translate into lower forecast error; the mechanism is signal attenuation.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The synthetic datasets accurately isolate and represent the non-stationarity types (trend, seasonality, heteroscedasticity) that matter for real forecasting performance, and the consensus ratio from ten statistical tests reliably identifies when a transform is matched to the data.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Large-scale experiments on synthetic data find stationarity transformations improve forecasts in only 18% of matched cases, with variance stabilization as the main exception and signal attenuation as the mechanism.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Stationarity transformations improve time series forecasts only 18 percent of the time even when matched to the data.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"45cd6a8d98adfa380e53767b7b10b11df245f3a2dd78513cc6871a03106f610a"},"source":{"id":"2605.17689","kind":"arxiv","version":1},"verdict":{"id":"ffc82a68-1974-4e6c-838c-7738d996a9e4","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T22:02:52.986390Z","strongest_claim":"For matched pairs, transforms improve forecasts only 18% of the time. The primary exception is variance stabilization: log and Box-Cox on heteroscedastic data improve accuracy in 60-65% of cases. Mediation analysis confirms that while transforms achieve trend stationarity, this does not translate into lower forecast error; the mechanism is signal attenuation.","one_line_summary":"Large-scale experiments on synthetic data find stationarity transformations improve forecasts in only 18% of matched cases, with variance stabilization as the main exception and signal attenuation as the mechanism.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The synthetic datasets accurately isolate and represent the non-stationarity types (trend, seasonality, heteroscedasticity) that matter for real forecasting performance, and the consensus ratio from ten statistical tests reliably identifies when a transform is matched to the data.","pith_extraction_headline":"Stationarity transformations improve time series forecasts only 18 percent of the time even when matched to the data."},"integrity":{"clean":false,"summary":{"advisory":0,"critical":1,"by_detector":{"doi_compliance":{"total":1,"advisory":0,"critical":1,"informational":0}},"informational":0},"endpoint":"/pith/2605.17689/integrity.json","findings":[{"note":"Identifier '10.3390/1010000' is syntactically valid but the DOI registry (doi.org) returned 404, and Crossref / OpenAlex / internal corpus also have no record. The cited work could not be located through any authoritative source.","detector":"doi_compliance","severity":"critical","ref_index":19,"audited_at":"2026-05-19T22:11:05.062399Z","detected_doi":"10.3390/1010000","finding_type":"unresolvable_identifier","verdict_class":"cross_source","detected_arxiv_id":null}],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T22:31:19.427167Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T22:11:05.062399Z","status":"completed","version":"1.0.0","findings_count":1},{"name":"cited_work_retraction","ran_at":"2026-05-19T21:51:57.840035Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"shingle_duplication","ran_at":"2026-05-19T21:49:43.950933Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"citation_quote_validity","ran_at":"2026-05-19T21:49:43.749945Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T21:33:23.523001Z","status":"skipped","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T21:21:57.433528Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"753bf69e56dced24c61c54bdac7b004409fca36ebd7604f051c73ffe41b68a17"},"references":{"count":19,"sample":[{"doi":"","year":1970,"title":"Box, G.E.P .; Jenkins, G.M.Time Series Analysis: Forecasting and Control; Holden-Day: San Francisco, CA, USA, 1970","work_id":"0661c3fb-2960-4db1-a275-d0d25bcebc92","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Hyndman, R.J.; Athanasopoulos, G.Forecasting: Principles and Practice, 2nd ed.; OTexts: Melbourne, Australia, 2018","work_id":"87b62e56-7c95-4201-af68-f2b8fd7f02db","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"An analysis of transformations.J","work_id":"45aa2d19-2b48-4801-ad7b-f0e63dde7245","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1983,"title":"The estimation and application of long memory time series models.J","work_id":"8c468aca-03f3-4748-a6e4-8e859933273c","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2000,"title":"The M3-Competition: Results, conclusions and implications.Int","work_id":"f54739eb-2b27-4f02-b047-13c9a42a78d0","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":19,"snapshot_sha256":"59fa5e0387893fbe80f1f2e7096d8f662b5d6e3103f4cbf32130afc37ece2c7d","internal_anchors":1},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}