a Qjj=@s|dZddlZddlZddlZddlmZejddejddddl Z ddl Z dZ d dd Zd d Zed krxedS)aAQueue-aware fill model: what happens once you stop ignoring the orders ahead? WHY. Every fill number this project has produced -- historical replays, the live paper maker, the 0-5c price ladder -- uses the same test: "did any trade print at or above our ask?" That ignores QUEUE POSITION. In a FIFO book, a buyer lifting the ask takes the size already resting there first. With 1,749-44,427 contracts ahead of us at the touch (measured on Kalshi, 2026-07-28 onward), a print at our price is very often somebody else's fill, not ours. The bias is DIRECTIONAL, which is what makes it dangerous rather than merely imprecise: - at +0c we join the touch, so the queue ahead is huge -> fills badly inflated - at +3c/+5c we are the only order at that price -> fills roughly honest So the measured fill curve (steep decline from 100%) is flattered at the bottom and fair at the top. The TRUE curve must be flatter, which changes the answer to "should we charge more" and inflates the concede-at-T-60 policy specifically, since conceding puts us exactly where the inflation is worst. THE MODEL. When we post at price P with Q contracts already resting at P: filled once cumulative volume of trades at price >= P (after we post) exceeds Q. Trades at a HIGHER price than ours mean the book walked up through our level, so they count too. Q is taken as the depth at the touch at POST time when we join the touch (our_ask <= market ask), and 0 when we post above the touch (we are first in line at a price nobody else is showing). KNOWN LIMITATIONS, stated rather than buried: - Q is pessimistic: makers ahead of us can CANCEL, shrinking the real queue. So this is a lower bound on fills, as the optimistic model is an upper bound. The truth is between them, and the gap is the honest uncertainty. - We only have depth at the BEST ask, not the full ladder. For orders above the touch we assume nothing is queued at our level, which is optimistic there. - Depth logging only began 2026-07-28 17:17Z, so earlier legs cannot be scored. Usage: python research\queue_aware_fill.py [YYYY-MM-DD] N) defaultdictzE:\soccer\mainzE:\soccer\researchz&E:\soccer\research\mm_paper_quotes.csv@c CsFgd}}t||kr|dd}|r,||d<ztd|}WntyRYqYn0|dg}||7}|d}|r|s|qtdd|Ddd }|r |r ||kr qq g} |D]} t| d } | r|r| |krqz>| | t | d pd t | d p| dpd fWqt t fy.YqYq0q| j ddd| S)z>[(ts, price, count)] after `since`, paginated back far enough.N)tickerlimitcursorz/markets/tradesZtradescss.|]&}t|drt|dVqdS) created_timeN)S_ptget).0xr&E:\soccer\research\queue_aware_fill.py Dsztrades_sized..)defaultrZyes_price_dollarsrZcount_fpcountcSs|dS)Nrr)zrrrWztrades_sized..key) lenktZ_get Exceptionr minr r appendfloat TypeError ValueErrorsort) rZsinceZcapoutrpdtrZoldestresr trrr trades_sized4s@         r'csttjdkrtjdnd}tt}ttddZ}t|D]>}|ddd|kr<| dr<| dr<||d  |qt d?t d@dS)ANz 2026-07-29zutf-8)encodingts Zour_askZk_askZ leg_tickerz: z Kalshi legs with quotes cSsdddgdS)Nr)noptqqueuesrrrrrrdrzmain..cSs|dS)Nr*r)r rrrrfrrrZask_sizeZkickoffd& .>gcs$g|]\}}}|kr|||fqSrr)r r&r"c)korr rrzmain..c3s |]\}}}|dkVqdS)r2Nr)r _r")ourrrrsrzmain..FTr,r-r.r/z4FILL RATE: optimistic (ignores queue) vs queue-awarez armz>5s Zlegs optimisticz>11sz queue-awarez>12sz median queuez>13sz +zc Z5d/z6.0fz% z7.0fz12.0fz TOTAL Z3dz legs: optimistic z (z.0fz%) queue-aware z%)z. => the optimistic model overstates fills by z percentage pointszA Optimistic is an UPPER bound; queue-aware is a LOWER bound (itz: assumes nobody ahead of us cancels). Truth lies between.)rsysargvrlistopenQUOTEScsvZ DictReaderr rprintitemsr stripr r rroundr'anysorted)ZdayZrowsZfhrZper_armZtkZrsZr0startZmktr8Zqueuer$r:ZcumZ queued_fillr6r"r3aZtntoZtqr,Zmedr)r4r7rmain[s(2      $  6 (  rM__main__)r)__doc__rBZdatetimeZdtr= collectionsrpathinsertZ mm_settler Z kalshi_traderrrAr'rM__name__rrrrs&  '>