a Xj<@sdZddlZddlZddlZddlmZddlmZdZgdZddZ d d Z d d Z d dZ ddZ ddZdddZddZddZddZddZedkredS)ai Does the RAW SIZE of the Kalshi-vs-Poly CLV gap predict profit, with NO band gate at all (i.e. bet the cheaper book regardless of whether either leg clears [5,12) vs Pinnacle)? NO. Distinct from two prior negative results this complements: - kalshi_poly_pure_deviation.py: fair = mean(kalshi,poly), no Pinnacle at all -> dead. - forward_test_unique_book.py fire_xlow(): the LIVE rule, which REQUIRES the cheap leg's clv-vs-Pinnacle to land in [5,12) AND the other leg to sit OUTSIDE that band. This script asks the question in between: keep Pinnacle as the reference for each leg's CLV (so it's not "pure" deviation), but drop the band gate entirely and ask whether the raw GAP SIZE between clv_kalshi and clv_poly (both vs pin fair) is itself predictive, continuously, regardless of where either value falls. METHOD: first pre-KO snapshot per fixture; for each side, cheap leg = lower- priced book, gap = clv(cheap leg) - clv(other leg) [always >= 0]; pick the side with the single largest gap; settle at the cheap book's ask (Kalshi ask / Poly est-ask). Bet UNCONDITIONALLY -- no band requirement anywhere. RESULT on poly_backfill (643 fixtures with a resolvable K/P gap, 2026-07-16): Disjoint gap buckets are NOT monotonic (the tell something's wrong): [0,2)% n=305 ROI -4.0% z-1.05 [2,5)% n=255 ROI +4.2% z-0.01 [5,8)% n= 50 ROI -24.1% z-0.86 <- dips hard, breaks monotonicity [8,12)% n= 22 ROI -8.6% z+0.86 [12,20)% n= 11 ROI +77.5% z+1.37 <- only n=11, ~7% implied (longshot) Cumulative thresholds all collapse to a handful of ultra-longshot bets as T rises (n: 643->83->33->8 as T: 0%->5%->8%->15%); the ONE "significant" cell (gap>=15%, n=8, ROI+144%, z+2.06, p=.02) is 8 bets at ~7% implied odds -- one win at ~14:1 explains the whole result. Classic small-cell longshot mirage, same shape as the round-robin/sub-band traps already catalogued. OOS 60/40 split KILLS every threshold tried: gap>=0%: TRAIN -8.2% -> TEST +9.6% (sign flip, both near zero -- just noise) gap>=5%: TRAIN +5.6% -> TEST -23.9% (sign flip) gap>=8%: TRAIN -12.9% -> TEST +64.9% (sign flip, both small n) No threshold holds sign across the split. VERDICT: raw cross-book CLV-gap magnitude carries NO exploitable edge without the band gate. This is not a contradiction of the frozen strict-xlow rule -- it's confirmation that the band conditions (cheap leg IN [5,12), other leg OUT of [5,12)) are doing real selection work, not just being an arbitrary cutoff. Don't drop or loosen the band based on gap size alone. Read-only, backtest only. Mirrors kalshi_poly_pure_deviation.py conventions. N)datetime) defaultdictzE:\soccer\poly_backfill)z%Y-%m-%d %H:%M:%Sz%m/%d/%Y %H:%Mz%m/%d/%Y %H:%M:%Sz%Y-%m-%d %H:%Mc Cs8tD].}zt||WSttfy0Yq0qdSN)_FMTSrstrptime ValueError TypeError)sfr *E:\soccer\research\kalshi_poly_gap_size.py_pts9sr c Cs(z t|WSttfy"YdS0dSr)floatrr)xr r r _f>s rcCs8t|D]*}|dpd}|dvr|SqdS)Nftr)HDA)reversedgetupperstrip)rowsrvr r r _resultAs rcs(dd|D}t|fdd|DS)NcSsg|] }d|qS)?r ).0rr r r Gzfair_mult..csg|] }|qSr r )rZrir r r r Gr!)sum)orr r"r fair_multFsr%cCs(|dkr dS|dkrdS|dkr$dSdS)N皙?g~jth?g?g{Gzt?gffffff?g?{Gz?r )midr r r half_spreadHs   r)c Cstt}ttdD]h}t|dddF}t|D]*}||d|d|df|q6Wdq1sv0Yqg}| D]^\}}|d}|sqt d d |Dd d d }t |} |r| r|t || dd |Ddq|j dd d |S)Nz\*_polybackfill.csvrzutf-8)newlineencodingZ game_datehomeawayrcSs>g|]6}|dpddkrt|drt|d|fqS)ZstatusrZPREZ timestamp)rrr )rrr r r r Xs"zload..cSs|dS)Nrr )tr r r Zr!zload..keycSsg|] \}}|qSr r )r_rr r r r ]r!)gdrprecSs|dSNr3r )r r r r r/^r!)rlistglobDIRopencsvZ DictReaderrappenditemssortedrdictsort) ZfixturespathZfhroutr1rr3r4rr r r loadNs"H rB)r,r-c Cs|dD]}t|dt|dt|d}}}d|||fvs|dkr|dkr|dksdqt|||f}g}tD]d\}}t|d|d} t|d|d } t|d |d } | dus~| dus~d | krdkr~nq~d | kr dksq~q~| | krP| | r@d | kr8dkr@nn| n| | } } }n| | t| | } } }||| ||}|||||}||}|d dkr|dkp|d dko|dk}|t|||| ||ddq~|rt|dddSqdS)zFirst pre-KO snapshot: pick the side/book with the LARGEST clv-vs-Pinnacle gap between Kalshi and Poly. No band requirement anywhere. Bet unconditionally.r4Zpin_homeZpin_drawZpin_awayNZkalshi_Z_midZ_askZpoly_Z_pricerrrr,rr-r3)clvgapsidepxwonr3cSs|dS)NrFr )dr r r r/ur!zpick_no_band..r0) rrr%SIDXr<r)r;r>max)ZfxrZphZpd_ZpaZpfZcandsrGZsiZkmZkaZpmr(rHotherrEZoclvrFrIr r r pick_no_bandcs".(,DB(. rNcCst|}|sdStdd|D}tdd|D}tdd|D}tdd|D}|dkrr||t|nd}t||||d||d||d|d t|td d S) Ncss|]}|dVqdS)rINr rbr r r {r!zstat..css.|]&}|dr"d|d|dndVqdS)rIrDrHgNr rOr r r rQ|r!css|]}|dVqdS)rHNr rOr r r rQ}r!css"|]}|dd|dVqdS)rHrDNr rOr r r rQ}r!rdg?rC)nwwriwrroizp)lenr#mathZsqrtr>Zerfc)betsrTrUZnetZmuvarrYr r r statxs$&r_cCs|std|dddStd|dd|ddd|dd|d|dd |d d d |d d d|ddd|ddd|dddS)Nz Z26sz (none)z n=rTZ3drUzW-zL WR=rVz4.1fz% (impl rWz%) ROI=rXz+7.1fz% z=rYz+.2fz p=rZz.4f)print)tagr r r r lines@rbcs`t}dddd|DD}|jdddtdt|d td d D]D\fd d|D}td ddddddt|qNtddD]2fdd|D}tddddt|qtddD]ztfdd|Dddd}tt|d}tddddt|d|tddddt||dqdS)NcSsg|] }|r|qSr r rOr r r r r!zmain..css|]}t|VqdSr)rN)rr r r r rQr!zmain..cSs|dSr5r rPr r r r/r!zmain..r0z4Total fixtures with a K/P gap on at least one side:  z$=== Disjoint buckets by GAP SIZE ===))rr')r'r&)r&{Gz?)reQ?)rf皙?)rgrcs,g|]$}|dkr krnq|qSrFr rO)hilor r r r!zgap [rSz.0f,z)%z< === Cumulative gap thresholds (bet if gap >= T, no cap) ===) rRr'gQ?r&reg?rfg333333?rgcsg|]}|dkr|qSrhr rOTr r r r!zgap >= %z === OOS 60/40 check ===)rRr&recsg|]}|dkr|qSrhr rOrlr r r r!cSs|dSr5r rcr r r r/r!g333333?zT=z% TRAINz% TEST)rBr?r`r[rbr_r=int)Fr]subZcutr )rmrirjr mains$ *$rr__main__)__doc__r:r7r\r collectionsrr8rr rrr%r)rBrKrNr_rbrr__name__r r r r s$1