a xj0&@sdZddlZddlZddlZddlmZejddddlZ e j Z e j Z ddZ ddZ dd d Zdd dZdddZddZddZedkredS)aCandidate-rule sweep with an explicit train/test split. Trades nothing. WHY THESE CANDIDATES AND NOT A GRID SEARCH. With 1,717 fixtures and a flexible rule space, something will always backtest well -- `unique` shows +20.6% at n=55 and is a hindsight artifact whose executable cousin t60 runs -18.3%. So the candidates here are limited to three MECHANISMS, each motivated by a measured finding rather than by scanning for a winner: 1. PRICE FLOOR. The binding problem is variance, not edge: xlow's per-bet stdev is 1.94u at a mean price of 0.26, so proving its observed +0.8% ROI needs ~100,000 bets (~46 years). Variance of a unit bet scales as (1-p)/p, so restricting to higher-priced outcomes cuts the required n roughly 4x at p=0.5 versus p=0.26. This does not create edge; it makes edge MEASURABLE. 2. PERSISTENCE. forward_test_unique_book.fire_xlow's own docstring flags this as "a candidate refinement for the n>=100 review -- NOT wired now": on a 1c grid with flip-flopping books, scanning for the first qualifying snapshot finds TRANSIENT configurations. Requiring the shape to hold N consecutive snapshots should drop the noise fires. Pre-registered, not dredged. 3. KALSHI-ONLY. xlow needs both books and only 411 of 1,717 fixtures carry both; 1,364 carry Kalshi+Pinnacle. A single-book rule has 3x the sample, which matters more than rule elegance when the whole problem is power. DISCIPLINE. Train = on/before FREEZE_DATE, test = after. Report BOTH always; a candidate that only works in train is dead. Note that xlow's core was itself fitted on train, so train numbers for its variants are optimistic by construction -- only the TEST column is evidence, and only the INCREMENT over baseline xlow is attributable to the new dimension. MULTIPLE COMPARISONS. Every candidate tried is printed, including losers, with the count. At ~20 candidates, one crossing p=0.05 by luck is the expectation, not a finding. Usage: python research\rule_sweep.py N) defaultdictzE:\soccer\researchcCsnt|dt|dt|d}}}d|||fvsZ|dkrZ|dkrZ|dks^dSt|||fS)NZpin_homeZpin_drawZpin_away)FTZ_fgetZ fair_mult)rZphZpd_Zpar E:\soccer\research\rule_sweep.py _pin_fair1s4&r c Cst|||\}}t|||\}}|dus<|dus<|dur@dSd|krTdkrnnnd|krldksrndS||kr~dS|t|} | dkrdS| || } | || } td| krtdksndStd| krtdkrnndS| ||fS)zD(clv, mid, ask) if `book` is the exclusive low on `side`, else None.Nrr)r book_quoteSIDXBAND) rpfsidebookothermaZom_fairclvZoclvrrr _qualifies8s$.    rrc CsPg}|D]@}tt}d}|dD]}t|}|dur8q"i} dD]>} dD]4\} } t||| | | } | rH| d|krH| | | | f<qHq@t|D]}|| vrd||<qd}| D]>\}} ||d7<|||kr|dur|d|d| f}q|r"|}qq"|sq|\} } \}}}|d| dkr"d nd k}||r>d |d nd ||fq|S) zOne bet per fixture: first pre-KO snapshot where the shape holds for `persist` CONSECUTIVE snapshots. Returns list of (ret, ask, won).NprehomeZaway))kalshipoly)rrrrftrrHA?)rintr rlistitemsappend)fixtures min_pricepersistoutfxstreakpickedrr liverrrqkhitrZmidaskwonrrrrun_xlowNs>     $r3cCsg}|D]}tt}d}|dD]}t|}|dur:q"i} dD]} t|d| \} } | dusB| dusBd| kr|dksqBqB|t| } | dkrqB| | | }td|krtdkrBnqB| |krB| | | <qBt|D]}|| vrd||<qd}| D]4\}} ||d7<|||kr|dur|| f}q|r"|}qDq"|sLq|\} }|d| dkrhd nd k}| |rd |d nd ||fq|S) zIKalshi mid-CLV in band vs Pinnacle. No second book required -> 3x sample.Nrrrrrrrrrr r! rr"r rr r r r#r$r%)r&r'r(r)r*r+r,rr r-rrrrrr/r0r1r2rrrrun_kalshi_onlytsH &  (    $r5cCsg}|D]}tt}d}|dD]<}t|} | dur:q"i} dD]} t||| \} } | dusB| dusBd| kr|dkrBnqBd| krdksqBqB| t| }|dkrqB|| |}td|krtdkrBnqB| |krB| | | <qBt|D]}|| vrd||<qd}| D]:\}} ||d7<|||kr|dur|| f}q|r"|}qbq"|sjq|\} }|d| dkrdnd k}| |rd |d nd ||fq|S) awSingle-book mid-CLV in band vs Pinnacle, for any one book. Poly-US matters here for two reasons beyond sample size: it pays a MAKER REBATE (0.0125 x p x (1-p)) where Kalshi charges a taker fee, and it is the only venue covering GAP_LEAGUES. Note its history only starts 2026-07-24, so it has no pre-freeze period -- the split for it is internal by date. Nrrrrrrrrr r!r4)r&rr'r(r)r*r+r,rr r-rrrrrr/r0r1r2rrr run_book_onlysH >  (    $r6c Cst|}|dkrdStdd|D}tdd|Dp>d}|t|}t||||tdd|D|tdd|D|d |tt |d d d S) Nrcss|]}|dVqdS)rNr.0xrrr zsummarise..cSsg|] }|dqS)rrr7rrr r;zsummarise..g& .>css|]}|dVqdS)Nrr7rrrr:r;css|]}|dVqdS)rNrr7rrrr:r;\(\?gư>r=)nroizwrimpsdneed) len statisticsZmeanZpstdevmathZsqrtdictsummaxabs)Zrsr?rrDZserrr summarises rMc Csdt}dd|D}dd|D}tt|dtjdt|dt|tdtdg}d D]4}d D]*}|d |d d |||fddfqpqhdD]4}dD]*}|d|d d |||fddfqqtddddddddddddddddddddddd g}|D]\}}t||t||} } | r&| sXq&||| | ft|dd| d!d"dd#| dd$d%| dd&d| d!d'dd#| dd$d%| dd&d| d(d)q&td*t|d+d,t|d-d.d/d|D} td0t| | D]@\}} } td|d1| d!d2d#| dd3d4| dd5qdS)6NcSsg|]}|dtjkr|qSgdr FREEZE_DATEr8frrrr<r;zmain..cSsg|]}|dtjkr|qSrNrPrRrrrr<r;z fixtures | train (<=z) z | test zGtrain numbers for xlow variants are OPTIMISTIC -- xlow was fitted here.z"only the TEST column is evidence. )r皙?g?333333?皙?)rr=z xlow minP>=z.2fz persist=cSs t|||SN)r3rSmppzrrrr;zmain..)rrTrUrV)rr=zkalshi-only minP>=cSs t|||SrX)r5rYrrrr\r; candidateZ36s zTRAIN nz>8sr@z>7srAz>6sz zTEST nzneed nz>9sr?Z8ddz+6.1fz% z+6.2fZ7drEz9,.0f z candidates tried. At p=0.05, ~g?z.1fz0 are expected to look significant by luck alone.cSs g|]}|dddkr|qS)r=rAr>r)r8rrrrr<r;zcandidates with TEST z > 1.96: z : test n=z roi=z+.1fz% z=z+.2f)rZ load_fixturesprintrFrQr%rM) r*ZtraintestZcandsrZr[ZrowsnamefnrbZgoodrrrmains\,  "  2  $rf__main__)rr)rr)rr)__doc__rHrGsys collectionsrpathinsertZforward_test_unique_bookrr r r rr3r5r6rMrf__name__rrrrs"$  & ( . %