a L$jD'@s dZddlZddlZddlZddlZddlmZej j ddde de dd d \Z Zd d gZgd ZdddZdZdZddedDZejeddZejedddded<eedZeeddkeddk@ZgZeD]Zd eZeeee keeek@Ze dj!ed!d"Zeed#<eeed$<ed%ed&ed'<ed(eed)<e"eqejeddZ#e#j d$d*d+j!ed!d"Z#e# dj$dd,Z#e%e#d'dkd-e#d'ej&e#d.<e#d/e#d#'ek(e)e#d0<e#j*gd1d2j$dd,Z#e%e#d0d3ke#d.d3d4e#d5<e+e#Z,e#d6-Z.e/d7e,d8e.d9e#j01d:d;e#j21d<d= e/d>e/d?e#3d6Z4e5e46e4d07e4d01e4d51e4d57d@Z8e#d01Z9ej:e8dAe9dBe8dCdDZ;e9d3e9ej:d-e8dCe8dCdDZe=e9d3e9e>d3d-Z?e9e?d3e9e?Z@ZAe/dFe@dGdHeAdGdIe9d:dJe?dKdL e@e8dMe@eAe8dCe8dN<ejBCdOe@e8dMeAe8dCe8dMe8dP<ejBCdQe@e8dMeAe8dCe8dMe8dR<e8 dNZDe/dSd6dTdUdCdVdUdWdXdUdYdZdUd[d\ eDED]R\ZFZGe/dSeFdTdUe)eGjHdVdUeGjId]dUeGjJd^d_eGjKd`daeGjLd`db q P(true edge < 0) B. Benjamini-Hochberg FDR -- per-league test vs breakeven, corrected for testing many leagues at once. C. Minimum-n eligibility gate -- a league isn't judged until it has >= MIN_N bets. D. Walk-forward validation of "drop leagues below an ROI threshold": tune the drop-set on TRAIN, lock it, score on the untouched TEST. Plus the key diagnostic: does TRAIN league-ROI even predict TEST league-ROI? Pure analysis -- nothing here touches the live pipeline. N)statszutf-8replace)encodingerrorsz E:\soccerz display.width)皙?gQ?homeaway)Z game_daterr HA)rr (g333333?cCs,g|]$}|dr|dkrtd|qS)z.csvztest.csvz live_data/)endswithpdZread_csv).0fr-E:\soccer\research\analysis\_league_honest.py !srZ live_dataT)Z ignore_indexZ timestampZmixedZcoerce)formatrZtsZstatusZPREelapsedZ clv_kalshi_first)ZkeepsideZclvZkalshi_Z_askZaskZ pin_fair_p_winF)Z ascending)dropg?oddsftrwon)rrr)Zsubsetgnetleaguezreconstructed z bets across z leagues | overall WR z.1%z | ROI z+.1% zn==============================================================================================================zXA1 EMPIRICAL-BAYES SHRINKAGE on WIN RATE (beta-binomial prior fit by method of moments))nwinswrroirr#r!)Zweightsgư>z prior: Beta(a0=z.1fz, b0=z) -> prior mean z , strength ~z.0fz pseudo-betsr" wr_shrunkrwr_logffffff?wr_hiz z<8 z>4zraw WRz>7z shrunk WRz>10z90% CIz>18z>7.1%z>10.1%z [z>5.1%,]zo ==============================================================================================================z]A2 EMPIRICAL-BAYES SHRINKAGE on per-bet NET ROI (James-Stein normal model) -> P(true edge<0)ccs8|]0}tjtj|kdftj|dfdVqdS)rr$r%N)siglocrpersum)rlgrrr Vr1s2r$Zddofw roi_shrunkg-q=Zpost_sdp_negz pooled within-league sd(net)=z.3fz | between-league sd(edge)=z (zleagues DO differzCtau2=0 -> no detectable between-league signal; all shrink to global)zraw ROIz>8z shrunk ROIz>11zshrink wz>9z P(edge<0)z n/az>9.1%z>8.1%z>11.1%z>9.2fzSB BENJAMINI-HOCHBERG FDR -- per-league two-sided test vs breakeven (mean net = 0))rr!r$zp)columnsr;Zrankbh_qz testing z% leagues; flag = BH q<0.05 AND losingZROIr:zraw pzBH qz flagzLOSE*z>+7.2fz>8.4fz- --> survive FDR as genuinely unprofitable: ZNONEz%C MINIMUM-N ELIGIBILITY GATE (MIN_N=z*) -- below this, no power; pool the leaguez eligible for judgement: z (/z leagues, z bets)zW everything else is statistically indistinguishable from the pooled rate at current n.zPD WALK-FORWARD: tune a 'drop leagues below ROI threshold' rule on TRAIN (first z.0%z ), score TESTcCsRt|}|dkr tdtdfS||jddt|}|dtj|fS)Nr%nanrr4) lenfloatmeanstdnpsqrtrnormcdf)rr!r:rrrzpsrJrDsizete_roiinner)Zhowz1 D0 PERSISTENCE (leagues with >=5 TRAIN bets, n=z leagues):z+ corr(TRAIN ROI, TEST ROI): Pearson r=z+.2fz (p=z.2fz) | Spearman rho=zc --> if ~0, a league being bad in TRAIN tells you nothing about TEST, and dropping CANNOT work.z8 D0: too few overlapping leagues to assess persistence.z# TEST baseline (drop nothing): n=z>3z>+7.1%z | z=z | p=z.4fz7rule (drop if TRAIN ROI < thr & TRAIN league-n >= minL)z<52Zdroppedz<22ZTESTnz>5ZTESTroi)rrN )r5ggпgzthr=z(none)z>+8.1%z>7.4fz$ LOCKED rule (max TRAIN ROI): thr TEST: n=z vs baseline p=z@ Honest read: if the locked rule's TEST p is not clearly below z%, league-dropping is in-sample noise.)__doc__ossysZnumpyrFZpandasrZscipyrstdout reconfigurechdirZ set_optionZEDGE_LOZEDGE_HIZSIDESZGAMEZ SIDE_LETTERZMIN_NZSPLITlistdirZdfsconcatZdfZ to_datetimeZnotnacopyZpreZentriesrcqZ sort_valuesZdrop_duplicatesappendr,Z reset_indexwhererAmapZastypeintZdropnarBNZnuniqueKprintrrDrgroupbygZ DataFramerKr/r.mZaverageZvar_obsZsamp_varmaxZtau2_wrZkappaZa0Zb0ZbetaZppfZshowZiterrowsr0rr!r#r&r'r(ZmuindexZssZsigma2varZtau2rGrHrIrZisnar8Zpnr$r7r6Zrowsr-rsrEr:absr;ZfdrZmmZilocZcummin_r>flagtolistZsig_loseZeligZsoZcutZtrZterJZaggZtr_roirLjoinrenameZbothZpearsonrZprZ spearmanrZspvaluesZzbZpbZbestZminLZthrrZisinZkepttagZddZtr_keptZtroirrrrs        & $ "2  &* ,, 0L  ""6*  8F "  4: R $ : 20  "B   &0