{"top/10":{"cs_difference":{"rows":2388,"series":940,"metric":"rmse","improvement":-0.044255919963198664,"interval_95":[-0.11296668303800685,0.023633931339645937],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"gold_difference":{"rows":2388,"series":940,"metric":"rmse","improvement":-3.16954115381003,"interval_95":[-5.5276823945505225,-0.9187729557696231],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"xp_difference":{"rows":2388,"series":940,"metric":"rmse","improvement":0.1245177839768985,"interval_95":[-1.4146131720664354,1.6697311635107996],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"}},"top/15":{"cs_difference":{"rows":2371,"series":937,"metric":"rmse","improvement":0.03816003108319421,"interval_95":[-0.058045574652933134,0.13648019792282423],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"gold_difference":{"rows":2099,"series":908,"metric":"rmse","improvement":-3.851163693003855,"interval_95":[-8.750961092170325,0.6588717518023438],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"xp_difference":{"rows":2276,"series":931,"metric":"rmse","improvement":-0.4032173625203086,"interval_95":[-2.895182464618463,2.2219432195460502],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"}},"mid/10":{"cs_difference":{"rows":2388,"series":940,"metric":"rmse","improvement":0.0,"interval_95":[0.0,0.0],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"gold_difference":{"rows":2388,"series":940,"metric":"rmse","improvement":0.0,"interval_95":[0.0,0.0],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"xp_difference":{"rows":2388,"series":940,"metric":"rmse","improvement":0.0,"interval_95":[0.0,0.0],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"}},"mid/15":{"cs_difference":{"rows":2369,"series":938,"metric":"rmse","improvement":0.09395729380957363,"interval_95":[0.026555286175140948,0.16064959096803758],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"gold_difference":{"rows":2100,"series":907,"metric":"rmse","improvement":6.105905133940837,"interval_95":[3.3233564591362694,8.9614829661223],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"xp_difference":{"rows":2265,"series":926,"metric":"rmse","improvement":3.0745844982401422,"interval_95":[0.6429032066665779,5.4954238010414445],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"}},"bot_duo/10":{"bot_cs_difference":{"rows":2388,"series":940,"metric":"rmse","improvement":-0.0889145387700836,"interval_95":[-0.2002347652264046,0.024579698586557258],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"duo_gold_difference":{"rows":2388,"series":940,"metric":"rmse","improvement":-11.498302541899648,"interval_95":[-14.520890130409935,-8.51236391111083],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"duo_xp_difference":{"rows":2388,"series":940,"metric":"rmse","improvement":-5.060086170462682,"interval_95":[-6.778960703376108,-3.359274664608211],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"}},"bot_duo/15":{"bot_cs_difference":{"rows":2383,"series":940,"metric":"rmse","improvement":0.0,"interval_95":[0.0,0.0],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"duo_gold_difference":{"rows":2007,"series":897,"metric":"rmse","improvement":0.0,"interval_95":[0.0,0.0],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"},"duo_xp_difference":{"rows":2272,"series":929,"metric":"rmse","improvement":0.0,"interval_95":[0.0,0.0],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"}},"global":{"rows":2536,"series":979,"metric":"log_loss","improvement":0.0,"interval_95":[0.0,0.0],"positive_means":"candidate has lower error","scope":"Conditional fixed-model held-series performance difference, not coefficient confidence or post-selection coverage"}}