a ;<$j  @sdZddlZddlZddlZddlZejjddde de ddd \Z Z d d gZ d d gZgdZdddZddedDZejeddZejedddded<eedZeeddkeddk@ZeedeZgZe D]Zd eZeeee keee k@Zedjed!d"Zeed#<eeed$<ed%eed&<ed'ed(ed)<ed'ed*ed+<ed,eed-<e eqejeddZ!e!jd$d.d/jed!d"Z!e!dj"dd0Z!e#e!d+dkd1e!d+ej$e!d2<e!d3e!d#%ek&e'e!d4<e!j(gd5d6j"dd0Z!e#e!d4d7ke!d2d7d8e!d9<e!d+e!d)e!d:<d;d<Z)e*d=e+e!d>e!j,-d?d@e!j.-dAdBe!j./dCdD e*dEe*dFe*dGdHD]:\Z0Z1e)dIe0dJdKe1dJdLe!e!j2e0ke!j2e1k@qe*dMe)dNe!e!j2dOke!j2dPk@e)dQe!e!j2dPke!j2dRk@e*dSe*dTdUD]*\Z0Z1Z3e)e3e!e!j4e0ke!j4e1k@qhe*dVdWD]*\Z0Z1Z3e)e3e!e!j5e0ke!j5e1k@qe*dSe*dXe!e!j6e ke!j6e k@Z7e!e!j6dkZ8e!e!j6dkZ9e)dYe!e)dZe7e)d[e8e)d\e9e*d]e!j6/d^e+e!d_e*dSe*d`e!j(d:gd6Z:e*dae:j;<dbdcdde:j;-dbdcdee:j;=dfdbdcdge+e:dL dhD]0\Z0Z1Z3e)die3e:e:j;e0ke:j;e1k@qe*djdkD](Z>e)dle>dbdmdne:e:j;e>kqe*dSe*dodp\Z?Z@ZAdqdrZBddtduZCdveDdwdxfdyeDdzdxfd{eDdwd|d}fd~eDdwdd}fdeDdddfdeDdddffD]L\Z3ZEeCfieE\ZFZGe*de3ddeFddeFe?d7ddeGdqdS)a= Five strategy-improvement probes on the live Kalshi CLV strategy (BOOK=kalshi, band [5%,12%), home/away, both-sides dedupe). Reconstructs the signal set from live_data keeping extra columns (b365 CLV, kalshi bid) so cross-book + spread analyses are possible. Pure analysis -- nothing here changes the live pipeline. Nzutf-8replace)encodingerrorsz E:\soccerz display.width)皙?Q?homeawayZUSLZBSA)Z game_daterr HA)rr cCs,g|]$}|dr|dkrtd|qS)z.csvztest.csvz live_data/)endswithpdZread_csv).0fr+E:\soccer\research\analysis\_ideas_study.py srZ live_dataT)Z ignore_indexZ timestampZmixedZcoerce)formatrZtsZstatusZPREelapsedZleagueZ clv_kalshi_first)ZkeepsideclvZ clv_b365_clv_b365Zkalshi_Z_bidZbidZ_askaskZ pin_fair_p_winF)Z ascending)Zdrop?oddsftrwon)rrr)Zsubsetgnetspreadc Csnt|dkr"td|dddStd|ddt|dd|jdd |jd d |jd d dS)Nr z<22z n= 0z n=z>4z | WR >5.1% | flat ROI z>+7.1% | net z>+7.1fu)lenprintrmeanr sum)namesrrrline2s r-zreconstructed z settled signals | WR z.1%r$z+.1%r%z+.1fr&zL(should match the strategy.ipynb baseline of 365 / 26.3% / +18.6% / +67.8u) zF======================================================================z9#1 CLV SUB-BAND -- is the edge uniform across [5%,12%)?))rQ?)r.Q?)r/{Gz?)r0皙?)r1r[z.0%,)z --- coarse halves ---z[5%,8%)rr0z[8%,12%)rzG ======================================================================zH#4 FAVORITE vs LONGSHOT -- where does the edge live? (by decimal odds)))r@z fav (<2.0))r5@z mid (2-4))r6 @z long (4-8))r7geAz bomb (>8)z" --- by p_win (Pinnacle fair) ---))333333?zp<15%)r9?z15-25%)r:皙?z25-40%)r;rzp>=40%zN#2 CROSS-BOOK AGREEMENT -- require B365 CLV to agree (currently Kalshi-only)zall (Kalshi signal)zB365 CLV also in bandzB365 CLV > 0 (soft)z B365 CLV <= 0z (B365 CLV available on /z signals)zJ#3 SPREAD-QUALITY FILTER -- skip wide Kalshi bid-ask (spread eats edge)?z spread (prob pts): median dz.1fz | mean z | p90 g?z (n=))r{Gz?z<=2c)r>{Gz?z2-4c)r?r.z4-6c)r.rz>6czspread z8 --- cumulative: keep only signals with spread <= X ---)r>gQ?r?r.z spread <= z.0fczI#5 SIZING / SHRINKAGE -- final $ and MAX DRAWDOWN (variance) from $1000)g@@r:rcCs0|d}t||d||d}t||tS)Nrr8)maxminCAP)prmultbrrrrkfrachsrGr8c Cst}t}d}tD]\}}|dur4t||d}nD|jd||j|j} tt| |j||d}|dkrxt|t }||j r||jdn| 7}t||}t||||}q||fS)Nr8rr) BR0sigZiterrowsroundrrrGrrAFLOORr) rEshrinkflatbrZpeakZmdd_rZstakerDrrrsimjs rRzquarter-Kelly (base)r:)rEz eighth-Kellyg?zqKelly, 50% edge-shrinkg?)rErMzqKelly, 75% edge-shrinkg?zflat 1% of bankrollg{Gz?)rErNzflat 2% of bankrollr>r"z<26z final $z>8,.0fz (z>+6.1%z) | max drawdown r#)r8N)H__doc__ossysZnumpyZnpZpandasr stdout reconfigurechdirZ set_optionZEDGE_LOZEDGE_HIZSIDESZEXCLUDEZGAMEZ SIDE_LETTERlistdirZdfsconcatZdfZ to_datetimeZnotnacopyZpreZisinZentriesrr@qZ sort_valuesZdrop_duplicatesappendrJZ reset_indexwherenanmapZastypeintZdropnar-r(r'rr)r r*lohirZnmrrrZ both_bandZb365_posZb365_negZspr!ZmedianZquantileZthrrIrCrLrGrRdictkwrZddrrrrs        $ "> 4""    $L(&