a Kj@sdZddlZddlZddlZddlZdZdZddlZeej ddZ ddZ d d Z d d Zd dZddZddZedkredS)aDo legs on wildly-swinging games fill more often? The hypothesis is a good one and the fill model cannot answer it: of its 24 features not one measures price movement. It has spread, overround, volume, depth, time -- all snapshots of a moment, nothing about how much the price travels. The mechanism would be simple. A resting ask fills when a buyer crosses it. A leg whose fair value wanders re-tests our price repeatedly; a leg pinned at one number never does. So movement should buy fills. CIRCULARITY IS THE WHOLE DIFFICULTY. Measuring a leg's volatility over its full quoting life and correlating that with whether it filled is close to measuring the fill itself: the move that lifted us is inside the window. So volatility is measured over an EARLY window only -- the leg's first `EARLY_N` snapshots -- and the fill outcome is taken from the whole life. Legs whose price history is shorter than the early window are dropped rather than scored on a partial one. That still leaves a leg that filled inside the early window contaminating its own predictor, which is why the split is reported both ways: all legs, and legs that survived past the early window. Nz&E:\soccer\research\mm_paper_quotes.csvz'E:\soccer\research\mm_paper_results.csvEARLY_Nc CsHz*t|}d|krdkr&nn|ndWSttfyBYdS0dS)Nr)float TypeError ValueError)xvr (E:\soccer\research\mm_volatility_fill.py_f$s "r cCstt}tt}ttdddd}t|D]H}|dt|d|d}}}|r.|r.|r.|| ||fq.Wdn1s0Y| D]"\}}| dd |D||<q|S) z#{leg_ticker: [k_ask in time order]}utf-8replaceencodingerrors leg_tickerZk_asktsNcSsg|] \}}|qSr r ).0_ar r r 7zload_prices..) collections defaultdictlistopenQUOTEScsv DictReadergetr appenditemssort)seqrowsfhrtrrZrsr r r load_prices,s  & 2r)cCsi}ttdddb}t|D]F}|d}|s2qt|dp@ddv}||dp^t|||<qWd n1sz0Y|S) z{leg_ticker: filled 0/1}r rrrfilled0)1TruetruerN)rRESULTSrrr strstripint)outr&r'r(fr r r load_fills;s 8r5cCsb|dt}t|tkrdSt|}t|t|}tddt||ddD}|||dS)zAMovement in the early window: stdev, range and number of changes.Ncss&|]\}}t||dkrdVqdS)g& .>rN)abs)rrbr r r Orzvol_of..r)sdrangechanges)rlen statisticsZpstdevmaxminsumzip)pricesZearlyr9rngr;r r r vol_ofHs    rDc s:td|dtdddddddd d dd d g}|D]\}fd d|D}t|dkrnq@tdd|D}||t|d|t|ftd|ddt|dd|d dd|t|dd q@t|dkr6|dd|dd}} td|ddd|ddd| |dddS)Nz by :z Zbucketz>14 Zlegsz>6r*z>7z fill ratez>10cs,g|]$}|kr krnq|qSr r rr'hilometricr r rXrzreport..css|]}|dVqdSr*Nr rGr r r r8[rzreport..Y@z>9.1f%rz -> z vs z: z+.1fz percentage points)printr<r@r!) r%rKZlabelZbucketsZgotnameZselr4Zlo_rZhi_rr rHr reportSs(  <rTc stt}}g}|D]B\}}||vr,qt|}|s:q|||d<t||d<||q|sntddSdtdd|Dt|}tt|dtd|d d t d d|Dfd d dD}td|dddd|dddd|ddddt |ddtd|ddf|d|ddf|d|ddf|dddfgt |ddgd d!d |D}t|d"krtd#t|d$t d%d|Dfd&d dD} t |dd'd| ddf| d| ddf| d| ddf| dddfgdS)(Nr*lifez!no legs with enough price historyrNcss|]}|dVqdSrMr rGr r r r8trzmain..z legs with >= z& snapshots of history, base fill rate z.1frOcss|]}|dVqdSr9Nr rGr r r r8xrcs g|]}t|tqSr r2r<rp)sdsr r ryrzmain..)g?g?g?z early-window stdev quartiles: rdz.2fzc / rrPcr9z'price stdev over the first %d snapshotsz calmest 25%Z2ndZ3rd z wildest 25%r;znumber of price changes))rrz never moved)rz1-2)r^z3-5)r_cz6+cSs g|]}|dtdkr|qS)rUrP)rrGr r r rr<z/ === legs that survived past the early window (z0, so the measured move came BEFORE any fill) ===css|]}|dVqdSrVr rGr r r r8rcs g|]}t|tqSr rWrX)sr r rrzprice stdev (uncontaminated)) r)r5r"rDr<r!rRr@rsortedrT) r$Zfillsr%r(rBr baseqZ survivorsZqqr )rbrZr mainds^         rf__main__)__doc__rrr=sysrr/osr2environr rr r)r5rDrTrf__name__r r r r s   +