a sJj'(@sdZddlZddlZddlZddlZddlZddlZdZdZej ej ej e dZdZdZdd d Zdd d ZddZdddZddZedkreedS)ahWho are the PROFITABLE market makers in Polymarket soccer, and what do they do? (read-only; touches no account and places nothing) WHY. Our own maker is stranded 13.8% of the time against an 11.1% break-even, and the open question is whether assembling a 3-leg book is even the right construct. Polymarket settles on-chain, so unlike Kalshi -- where trades are anonymous and only AGGREGATE maker P&L is recoverable (research/mm_feasibility.py) -- individual makers can be named, ranked by realised P&L, and studied. Polymarket models a soccer match as THREE binary markets ("Will X win?", "...end in a draw?", "Will Y win?"), which is exactly our three legs, so the comparison is structural rather than analogical. HOW A MAKER IS IDENTIFIED. data-api /trades returns the TAKER side only by default; with takerOnly=false it returns BOTH sides of every fill. The set difference on (transactionHash, proxyWallet) is therefore the maker side. Verified 2026-09-14: 901 of 942 hashes resolve to clean 2-row pairs and 53% of rows classify as maker, which is the ~half you would expect. P&L. Per wallet per market: cash flow (BUY -size*px, SELL +size*px) plus the settlement value of the net position (outcomePrices is ["1","0"]-style and aligns with outcomeIndex). A negative net position is the complement leg and the same arithmetic holds. READ THE OUTPUT SCEPTICALLY: * SURVIVORSHIP. These are wallets that are currently large. Anyone who ran the same strategy into the ground is absent, which is exactly the inference error that produced the South America mirage. * NO QUEUE DATA. Fills are visible; resting orders are not. Nothing here can speak to queue position or refresh cadence. * DIFFERENT ECONOMICS. Polymarket's fee and tick structure is not Kalshi's, and our verdict rests on the 1c tick being the binding constraint. Findings transfer as HYPOTHESES. * SIZE. The top wallet moved 2.2M on one match. What works while warehousing that inventory need not work at $1/leg. * The offset cap on /trades is ~10k rows per market, so the very largest markets are TRUNCATED (reported per market as `truncated`). Nz https://gamma-api.polymarket.comzhttps://data-api.polymarket.comzpm_maker_study.csvii'c Cst|D]z}ztj||dd}Wn(tyHtd|dYqYn0|jdkr`|S|jdkrpdStd|dqdS)N)paramsZtimeoutg?)rangerequestsget ExceptiontimesleepZ status_codejson)ZurlrZtriesirrresearch/pm_maker_study.py_get7s     rTc Csg}dD]}ttdd|dt|ddd}|s:q|D]f}|dpNg}t|d kr>d |d pjd vr>td d|Dr>tdd|D}|||fq>t||d krqq|j ddddd|d|DS)zRzmatch_events..css |]}t|dpdVqdS)rrN)floatr r rrrr#Sr$cSs |d S)Nrr)ZkvrrrWr$zmatch_events..keycSsg|] \}}|qSrr)r!_errr Xr$z match_events..N) rGAMMAstrlowerr lenallsumappendsort)n_eventsroutoffevsr*ZmsZvolrrr match_eventsEs(  "r8cCszgd}}|tkrn|t|d}|s*d|d<ttd|}|s@qn||7}|t|7}t|tkrbqntdq ||tkfS)Nr)marketrrr takerOnly/tradesgQ?) OFFSET_CAPPAGErDATAr/r r )cidZ taker_onlyrowsr6pdrrrtrades[s     rCrc sZ|r t|nt|}tt|d|r(dnddtdd}d}|D]}||kr^qV|dpjdd d }|rd }|d D]4} tt d | ddt ddd} | rd}qq|rqJ|d7}td|ddddddd} |d D]H} | d} zddt | dDWnt y<YqYn0t| d \} }t| d\}}dd|D}| t| 7} tt}tdd}| D]b}|d}z@t|d pd}t|d!pd}t|d"pd}WnttfyYqYn0|d#|f|v}|r*||d$d7<n||d%d7<||d |7<||d&|d'||d(|d'| |d)pdd*kr||||8<||||7<n(||||7<||||8<q|D]@\}}tfd+d,||D}||d-||7<qqtd.| d/d0qJ|S)1aAggregate maker behaviour over matches with COMPLETE trade history. THE TRUNCATION PROBLEM. /trades refuses offset > ~10500, and before/after filters are ignored, so the biggest markets return only their most RECENT 10k rows per leg -- a slice pressed right up against settlement, where trades happen at 0.999/0.001 and are near-riskless cleanup rather than market making. Measuring "legs quoted per match" on that window biases hard toward 1, because a maker who worked all three legs during the match but only one at the end reads as a one-leg maker. So: walk DOWN the volume ranking and keep only matches whose every leg came back under the cap. Smaller markets, but the history is whole and the number means what it says. z+ candidate matches with the 3-leg structurez" (selecting complete-history ones)r cSsddddtttdS)Nr)pnl maker_fills taker_fillssizeeventslegs)set collections defaultdictrrrrr&s zstudy..rrasciireplaceFrr;Z conditionIdrrr)r9rrr:Tz N4z<52)endflushcSsg|] }t|qSr)r%)r!xrrrr+r$zstudy..rcSsh|]}|d|dfqS)transactionHash proxyWalletr)r!trrr r$zstudy..cSs ttS)N)rNrOr%rrrrr&r$rWrJZpriceZ outcomeIndexrVrHrIrKZslugrLZsideZBUYc3s.|]&\}}||tkr |ndVqdS)rFNr/)r!rqZpricesrrr#szstudy..rG z>6z rows)r8printr/rNrOr encodedecoderr>r<rloadsr rCr%int TypeError ValueErroraddupperitemsr1)r4Z complete_onlyZpoolr7WZusedr*rZ trunc_anyr"rBZn_rowsr?ZallrZtruncZtkr)ZtkeyZcashZposnrXwZszZpxZoiZis_makercZsettlerr\rstudylsz         ( *,  rkcCsttjdkrttjdnd}t|}g}|D]\}}|d|d}|dkrVq4|d|}dd|dD}||t|d d |t|d t|d t|d |rtt |t|d nd|rtt dd|Dt|d nddq4|j dddddl }t t ddddJ} |j| |r:t|dngdd} | | |Wdn1sn0Ytdt|dt ddd|D} tdtdd | dd!D]6} td"| d#| d | d$| d%| d | d&fq| rd'd| D} td(t| t| d)t| t| f| rtd*t d+d| Dt| }t d,d| Dt| }td-|td.|dS)/NrrDrHrIcSsg|] }t|qSrrZ)r!vrrrr+r$zmain..rLrGrrJrKrcss|]}|dkrdVqdS)rrNr)r!Lrrrr#r$zmain..)walletrGfills maker_sharerJrKavg_legs_per_eventall3_event_sharecSs |d S)NrGr)rrrrr&r$zmain..r'rirzutf-8)newlineencoding)Z fieldnamesz wrote z wallets -> rEcSsg|]}|ddkr|qS)rrg333333?rr!rrrrr+r$z*TOP MAKER-DOMINANT WALLETS BY REALISED P&Lz %-44s %10s %7s %6s %6s %s)rprGrqzmk%Zevtszlegs/evt z! %-44s %10.0f %7d %6.2f %6d %.2frprqrrrscSsg|]}|ddkr|qS)rGrrrwrrrr+r$z7 maker-dominant wallets: %d profitable: %d (%.0f%%)rz8 THE QUESTION: legs quoted per match, profitable makerscss|]}|dVqdS)rsNrrwrrrr#r$css|]}|dVqdS)rtNrrwrrrr#r$z< mean legs/match : %.2f (3.0 = full book like ours)z share of matches all3: %.2f)r/sysargvrbrkrgvaluesr2roundr1r3csvopenOUTZ DictWriterlistkeysZ writeheaderZ writerowsr^)nrhr@risZtotZmsharerLr}ZfhZw_ZmkrZprofaZs3rrrmainsZ   & 2      r__main__)r)T)rDTr)__doc__rNrosryr r r,r>pathjoindirnameabspath__file__rr=r<rr8rCrkr__name__exitrrrrs$'   T-