a ‰j2@sdZddlZddlZddlZddlZddlmZddlm Z e dZ dZ ddZ d d Z d d Zd dZdddddZddZddZddZddZedkredS)a&Is our Pinnacle feed too stale to anchor the strategy? Reads research/pin_fresh_capture.csv (real-time Pinnacle from pinnapi) against the API-Football relay we have used for everything to date, and answers four questions in increasing order of importance: 1. HOW STALE? update frequency of each feed. 2. HOW WRONG? distribution of |fresh_fair - stale_fair| in cents. This is the size of the error sitting inside every CLV number we have ever computed. 3. DOES THE SIGNAL SURVIVE? recompute the xlow band on BOTH feeds and compare which fixtures/sides qualify. Three outcomes matter: - signals that fire on both -> genuine, staleness was not driving them - fire on stale only -> ARTEFACTS. we were betting lag. - fire on fresh only -> edge we have been blind to 4. WHICH ANCHOR DOES KALSHI TRACK? for each side, does the Kalshi mid move toward fresh fair or stale fair over the following hour? If Kalshi tracks fresh Pinnacle, fresh is the true anchor and our band is mis-calibrated. Question 3 is the one that decides whether to pay for a feed. A large "fire on stale only" bucket means the historical edge was partly an artefact of measuring a live market against a 6-hour-old line. Usage: python research\pin_fresh_vs_stale.py N) defaultdict)Pathz(E:\soccer\research\pin_fresh_capture.csv)g?gQ?c Cs(z t|WSttfy"YdS0dSN)float TypeError ValueErrorxr (E:\soccer\research\pin_fresh_vs_stale.py_f$s r c CsPztjt|dd}Wnttfy4YdS0|jr@|S|jtjj dS)NZz+00:00)tzinfo) dtdatetimeZ fromisoformatstrreplacerrrtimezoneZutc)sdr r r _pt+s rcs>tdd|DsdSdd|D}t|fdd|DS)z9Proportional devig -> implied probabilities summing to 1.css|]}|o|dkVqdSNr .0r r r r 5zfair..NcSsg|] }d|qS)?r rr r r 7rzfair..csg|] }|qSr r rrr r r9r)allsum)orr rr fair3s r$cCsdtstdtdg}ttdd$}t|D]}tt|dt|dt|dg}tt|dt|d t|d g}t|d }|r6|durq6| ||d |d |df|d ||t|d| d | d| dd| dd| dd| dt| dt| dt| dt| ddq6Wdn1sV0Y|S)Nzno capture yet at z' -- run main\pin_fresh_capture.py firstzutf-8)encodingZfresh_hZfresh_dZfresh_aZstale_hZstale_dZstale_aZtslchomeawayZ kickoff_utcverdict xlow_fresh xlow_staleZ game_date k_home_mid k_away_mid p_home_mid p_away_mid)tkeyr&freshstaleZkor'r(r)r+r,gdr-r.r/r0) SRCexists SystemExitopencsvZ DictReaderr$r rappendget)rowsZfhr#frstr1r r r load<s0&&           *r@r-r.r/r0))r'kalshi)r(rA)r'poly)r(rBcCs||D]r}||pd}|rd|vr(q|dd\}}|t||fd}|rd|krfdkrnq||fSqdS)z~(side, price) of the FIRST snapshot where `col` fires, mirroring fire_xlow's one-bet-per-fixture, no-look-ahead selection.r*/rrN)r<stripsplit PRICE_COL)snapscolr#vsideZbookpr r r _first_bet_s rLcCs@|sdS|\}}|dkr |dkn|dk}|r8d||nddfS)zI(pnl, staked) in units for one bet, or (0, 0) if the anchor did not fire.)rMr'HArgrr )ZbetftrrJrKZwonr r r _retms rQc# Csz ddl}Wnty(tdYdS0i}|jddD]D}|drR|ddni}|d||d|d |d |d f<q:tt}|D]*}||d|d |d |d f|q|D]}|j d ddqgddgddgd}} tdd} d} | D]v\} } || }|dvr(q| d7} t | dt | d}}t ||\}}t ||\}}| dd|7<| dd|7<| dd|7<| dd|7<||||r|r| dd|7<| ddd7<| dd|7<| ddd7<n^|rN| dd|7<| ddd7<n.|r| dd|7<| ddd7<qtd| stddStd| d dD]f}| |\}}|dkrd!nd"}t|rd|d#|d$d%|d&d'||d(d)d* n d|d+qt |}t|}|dkrnNt|d?kr2td,|t|d@dnd} tdA| rNdB| dCdDndEtdFdGD]J}!| |!\}}"|"rbtdH|!dIdJ|"dKdL|d)d'||"d(d&d* qbdS)Ma4. DOES THE ROI DIFFER? -- the question the agreement stats cannot answer. Sections 1-3 measure how often the two feeds disagree. Disagreement is not a cost: a signal the stale feed invents is only expensive if it LOSES, and one it hides is only expensive if it WINS. This prices both. Run as PAIRED POLICIES over the same settled fixtures. Comparing two independent ROIs throws away the fact that both bet on largely the same matches; pairing removes fixture-level variance, which dominates everything else at this sample size. A fixture only one anchor fired on contributes that arm's P&L and zero to the other -- which is what running one policy and not the other actually does to your bankroll. PRICED AT MIDS, because mids are what the capture logs. That flatters both arms identically, so it cancels in the paired difference; the absolute ROIs below are NOT tradeable numbers. rNz; 4. ROI: forward_test_unique_book not importable -- skippedT)Zrequire_resultZprerPr5Zleaguer'r(r&cSs|dSNr1r rr r r rzroi_section..r2rM)r3r4cSsddgS)NrMrr r r r r rSr)rNDrOrr+r,r3r4both (fresh leg)both (stale leg)stale-only [ARTEFACT]fresh-only [MISSED]z: 4. DOES THE ROI DIFFER? (paired policies, priced at mids)zC no settled fixtures joined yet -- capture has no played matches. z settled fixtures comparedzfresh (pinnapi) zstale (API-Ftbl)  z4.0fz bets net z+8.2fzu ROI dz+7.2f%z no betsg\(\?z5 PAIRED per-fixture difference (fresh minus stale)z n=z mean=z+.4fzu se=z.4fz t=z+.2fz 95% CI [z+.3fz, z] units per fixturez differed on css|]}t|dkrdVqdS)& .>rNabs)rrr r r rrzroi_section..z of z3 fixtures -- the rest contribute nothing to the gapz2 SIGNIFICANT: the anchors do not pay the same.r^z9 NOT significant -- no measurable ROI difference yet.z' Resolving an effect this small needs ~,z fixtures.r*z where they disagree:)rVrWrXrYz Z24sz n=z<4dz net )Zforward_test_unique_book ImportErrorprintZ load_fixturesr<rlistr;valuessortitemsrLrQlen statisticsZmeanZstdevmathZsqrtr!r`int)#r=FresZfxZr0Zbyfixr#rIZdiffsZaggZbucketsZn_fixr2rGrPZbfZbsZpfZkfZpsZksnameZnetZstkZlblnZmuZsdZser1lohiZneedkZcntr r r roi_sectionvs  *(    (( (*    &$ .  rtcsRt}tt}|D]}||d|q|D]}|jdddq2tt|dt|dddd d fd d |DD}d d fdd |DD}td|rtdt |ddt |rddt |ddnd|rd?| td@|dAd=d>dBtdC|dDd=d>dEtdF|dGd=d>dHD]|}| |gddID]`\}}}} } tdJ|dK|dLt|ddMdNt|ddMdO| pdPdQ| pdP qq|dAd=}| dRkr6tdSd|| d+dTntdU| dVt|dS)WNr2cSs|dSrRr rr r r rSrzmain..rTz snapshots across z fixtures csvfdd|D}t|dkr"dS|dd|ddd}|dkrNdStdd t||d dD}||S) Ncs,g|]$}|rtdd|DqS)css|]}t|dVqdS)N)roundrr r r rrz0main..rate...)r<tuple)rr#rsr r rrz&main..rate..r1ricss|]\}}||krdVqdSrr rabr r r rrz%main..rate..r)riZ total_secondsr!zip)rIrsZvalsspanZchr rxr rates   zmain..ratecSsg|]}|dur|qSrr rr r r rrzmain..c3s|]}|dVqdS)r3Nr rrIrr r rrzmain..cSsg|]}|dur|qSrr rr r r rrc3s|]}|dVqdS)r4Nr rrr r rrz>1. UPDATE FREQUENCY (changes per hour, median across fixtures)z fresh Pinnacle (pinnapi) z.2fz/hz ~ every <z.0fz minr*z stale Pinnacle (API-Ftbl) z< (no stale odds captured -- API-Football returned nothing)r4cSs g|]\}}t||dqS)r\r_r{r r r rrr3zE 2. DISAGREEMENT between the two feeds (cents of implied probability)r\)rpz median zc p75 Jzc p90 Yzc p99 bzc max cz% share of legs disagreeing by >2c: css|]}|dkrdVqdS)rarNr rr r r rrz.1fr]zY For scale: the xlow band is 5-12% of fair, and a 25c leg in band is 1.25-3c from fair.z no paired rows yetz 3. WOULD THE SAME SIGNALS FIRE?r)rr+r,r&r'r(z? no scored snapshots yet -- capture has not written verdicts.cSs g|]}|dr|drdqS)ryrurr rr r r rrZsamerZz scored snapshotsz agree (same verdict) : rZ5dz of which BOTH fired: z stale-only [ARTEFACT] : stale_onlyz7 <- signals we bet that a fresh line does NOT supportz fresh-only [MISSED] : fresh_onlyz' <- edge the stale feed hides from usz differ (side/book) : differ)rrrz [z] r[z v z fresh=-z stale=z VERDICT: z/% of scored snapshots are stale-only artefacts.z Sample still small (z" snapshots); treat as directional.)r@rrer;rfrgrdrirjZmedianr~Z quantilesmaxr!rlr<rDrrt)r=Zbyfxr#rIr>srrqZverdZfiredZxfZxstotalZbothZbucketr&hr|Zartr rr mains   $& $ , " .  "4  r__main__)__doc__r:rrrkrj collectionsrpathlibrr6ZBANDr rr$r@rFrLrQrtr__name__r r r r s*    a[