[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-trimming-training-data-made-this-forecasting-ai-better":10,"sections":40},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":30,"tags":31,"sources":35,"feedback":39,"feedback_at":22,"cost_usd":39,"total_tokens":39},8626,"trimming-training-data-made-this-forecasting-ai-better","Trimming Training Data Made This Forecasting AI Better","A new arXiv paper (2609.37255) shows time-series forecasting models trained on carefully chosen data subsets outperform ones trained on the full dataset.","A new paper argues time-series forecasting models get better when they train on less data, not more, provided you pick the right less.\n\nResearchers behind arXiv:2609.37255, posted September 30, 2026, built a static data-selection framework for time-series foundation models, or TSFMs, the transformer-style models trained on huge, mixed collections of sequential data to forecast things like server load or retail demand. Instead of randomly sampling training windows, the team scores each window with a small reference forecaster and keeps only an intermediate difficulty range from every source dataset, on the theory that windows that are too easy or too hard waste training signal, while stratifying by dataset so no single source gets crowded out. Across multiple TSFM architectures, this loss-guided selection beat both random sampling and full-dataset training, as measured by MASE (mean absolute scaled error, a standard yardstick for point forecasts) and CRPS (continuous ranked probability score, which grades how well a model's uncertainty estimates match reality). The paper doesn't publish the actual score deltas, only that gains came \"by an absolute margin,\" which is vague framing for a fairly specific technical claim.\n\nThe bigger idea is that data curation, not just data volume, is becoming a real lever for foundation models, mirroring dataset-pruning and mixture-reweighting work already common in language-model training. What's specifically notable here is that a small, cheap reference model can select data for a much larger target model, provided the two models agree on which windows are relatively hard. That could cut the compute cost of building future TSFMs without touching model architecture at all.\n\nNone of this comes with a public leaderboard or independent replication yet, so file \"beats full-dataset training\" under promising, not proven.","[\"ai\",\"machine-learning\",\"time-series-forecasting\",\"research\"]","2026-09-30T04:00:00.000Z","2026-09-30T15:44:53.269Z","2026-09-30T15:44:59.480Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"Add the paper's identifying details (arXiv ID\u002Fdate, e.g. arXiv:2609.37255) since no source is named beyond 'a new paper,' and either drop the MASE\u002FCRPS mention or briefly explain what those metrics measure and give the actual comparison figures behind the 'beats full-dataset training' claim.","resolved","ai",[30,32,33,34],"machine-learning","time-series-forecasting","research",[36],{"name":37,"url":38},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.37255",0,{"sections":41},[42,45,49,53,58,63,68,73,78,82,87,92,97,102],{"name":43,"slug":30,"count":44,"latest_published_at":18},"AI",5146,{"name":46,"slug":47,"count":48,"latest_published_at":18},"Security","security",788,{"name":50,"slug":51,"count":52,"latest_published_at":18},"Policy","policy",417,{"name":54,"slug":55,"count":56,"latest_published_at":57},"Deals","deals",284,"2026-09-29T21:00:00.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":62},"Hardware","hardware",194,"2026-09-29T13:16:04.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":67},"Science","science",154,"2026-09-28T13:19:18.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":72},"Consumer Tech","consumer-tech",142,"2026-09-29T18:38:03.000Z",{"name":74,"slug":75,"count":76,"latest_published_at":77},"Software","software",91,"2026-09-25T20:55:00.000Z",{"name":79,"slug":80,"count":81,"latest_published_at":18},"Dev Tools","dev-tools",90,{"name":83,"slug":84,"count":85,"latest_published_at":86},"Startups","startups",83,"2026-09-29T21:51:36.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"General","general",49,"2026-09-28T16:44:57.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"Gaming","gaming",48,"2026-09-25T18:35:21.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":103,"slug":104,"count":105,"latest_published_at":106},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]