a tjZ3@sdZddlZddlZddlZddlZddlmZddlm Z ddl Z ddl m Z ddlmZmZddlmZddlmZe d Ze d Zd d Zd dZdddZddZddZdddZddZedkreedS)aCan we predict WHICH LEGS FILL, at the moment we post them? WHY. The maker's whole problem is partial fills: the prize is ~3c a book and every repair (crossing to complete, flattening a stranded leg) costs about that much. If fill probability is predictable at post time, the fix is to quote only books whose three legs are all likely to fill -- and skip the rest, which costs nothing. NO LOOK-AHEAD. Every feature is read from the FIRST quote row for that leg, so it is knowable at the instant we decide to post. Nothing about how the quote evolved, and nothing about the trade that filled it, is allowed in. THE BAR IT HAS TO CLEAR. We already know price distance predicts fills -- legs that walk to the touch fill ~25%, legs that stay 2c above fill 1-6%. A model that only rediscovers that is worth nothing. So the honest test is whether the full feature set beats a one-feature baseline on held-out data, not whether it beats random. Usage: python research\fill_model.py N) defaultdict)Path)LogisticRegression)StratifiedKFoldcross_val_score) make_pipeline)StandardScalerz&E:\soccer\research\mm_paper_quotes.csvz'E:\soccer\research\mm_paper_results.csvc Cs8zt|}||kr|ndWSttfy2YdS0dS)N)float TypeError ValueError)xvr E:\soccer\research\fill_model.py_f&s rc Cs>dD]4}ztjt||WSttfy6Yq0qdS)N)z%Y-%m-%dT%H:%M:%S.%f%zz%Y-%m-%dT%H:%M:%S%zz%Y-%m-%d %H:%M:%S)dtdatetimestrptimestrr r )sfrrr_pt.s rFc!CsBii}}tjddddz}t|D]^}|d}|s:q&t|ddvrTdnd ||<t|d pld t|d p|d ||<q&Wd n1s0Ytt d}t t |i}|jddddT}t|D]8}|d}|r||vr||vr|dsds|||<qWd n1s00Ygd}ggg} } } tttt} } |D]b\}}t|d}|d ur| |d|t|d}|d url| |d|ql|D]0\}}t|dt|dt|d}}}t|dt|d}}t|dt|d}}| |dp`g}d ||||fvs|r|d us|d urqd |kr|krdksnq||d}d |krdksnqd d l}t|d}t|d}t|d}t|d }| |dpDd g}| |||d!||d!|d"d!|d ur|d"d!nd t|t|t|dpd t|d#pd t|||d$krd"nd |d%krd"nd |d ur||d!nd t|d&pd t|d'pd t|d(p(d t|d)p:d t|d*pLd t|d+p^d |d,|j|prd d-|d,|j|pd d-|d ur||d!nd |d urd"nd g| ||| |d||||d t||d.pd/d0qt| tt| t}} |r8|| || fS|| |fS)1aB(X, y, names[, events]) -- one row per leg, features from its FIRST quote. ONE code path. An earlier version rebuilt the event list with a copy of the row filter, which can drift from this one and silently pair a leg's features with another leg's book -- the kind of bug that produces a plausible number. zutf-8replace)newlineencodingerrorsZ leg_tickerZfilled)1TruerZ pnl_centssizeNzmm_paper_quotes.snap.csvreason)our_askedge_over_mkt_cZspread_cZ overround_cZ over_mid_cZ min_leg_volZ book_vol_sumask_sizebid_sizeZbook_ask_size_minZ hrs_to_koZis_drawZis_homeZ fair_gap_carm mkt_age_h open_interest volume_24hno_bidno_askZhour_sinZhour_cosZpc_gap_cZhas_pcZvolumeeventr%r#Zk_bidZk_askZover_askZpin_fairtsZkickoffg @sideZover_midZpc_askZ local_hourY@?r&Zdrawhomer'r(r)r*r+r,g8@regime?)r-koZtickerpnlvolr5)RESULTSopencsvZ DictReadergetrstriprrtempfileZ gettempdirshutilZcopyfileQUOTESrlistitemsappendrZ total_secondsmathlowerminsumZsinpiZcosnpZarrayr int)! with_eventsZlabelZlegpnlZfhrttmpfirstZNAMESXyeventsZbookvolZbookaskr aZaskZkbZkaZoverZfairr.r7ZvolsZhrsrEr0ZomidZpcaZhourZasksZXaZyarrrbuild8s  "H  "*  .,&    rUcCs"t\}}}|j\}}t|d|dd|dd|dksb|dksb||dkrntddStd d d d }tttd dd}t ||||dd}| d}t ||dd|gf||dd} tddddddddtddd| dd| dtddd|dd| dtddd|| dtd t |D]F\} } t ||dd| gf||dd} td!| d"d| d#qh| |||d$jd } td%tt|| d&d'd(D]"\} }td!| d"d|d)qtd*t||dS)+Nz( legs with complete post-time features, z filled (r1z.0fz%) <z2too few events to fit anything honest -- stopping.Tr)Zn_splitsZshuffleZ random_stater2Zmax_iterCZroc_auc)cvZscoringr$ZmodelZ28 zCV AUC>8Zsd>7zprice distance onlyz>8.3fz>7.3fz all featureszlift from everything elsez>+8.3fz per-feature CV AUC on its own: Z20z.3fz8 standardised coefficients (full fit, sign = direction):cSst|d SNr)absrNrrrzmain..keyz+.3fz AUC 0.5 is a coin flip; 0.7+ is a usable signal. Judge the LIFT line, not the headline -- price distance is already known and already actionable.)rUshapeprintrHmeanrrrrrindexZstd enumeratefitZcoef_sortedzip book_readout)rQrRnamesnkr\pipefullZi_edgebaseiZnmrZcoefcrrrmains<   $   &&"   rzc Csddlm}tdd\}}}}t|dkr.dS|||||ddddd f}td d }t|||D]6\} } } || d d | || d d| qddd|D} t| dkrtdt| ddS| tdt| dtddddddddddddt d t| d } t d!D]\}}|d"krZ| || |d | n| d"| d}|stq0t t d#d|D}t t d$d|D}td|ddt|dd|d%d|d&q0td'dS)(a1A leg model is only useful if it separates COMPLETABLE BOOKS. A book pays only when all three legs fill, so the decision at post time is P(all three) -- the product of the per-leg probabilities. Probabilities are OUT-OF-FOLD, so no book is scored by a model that trained on its own legs. r)cross_val_predictTrLrVN predict_proba)r\methodrcSs ggdS)N)prRrrrrrrerfzbook_readout..r-rrRcSsDg|]<}t|ddkrt|dtt|dt|dfqS)rRr)lenrJprodrKall).0drrr sz book_readout.. only z+ complete 3-leg books -- too few to bucket.z( BOOK COMPLETION vs PREDICTED P(all 3), z books, out-of-foldr`ZbucketZ14r]booksz>6Zpredr_Zactualr^)z worst 25%Z2ndZ3rdzbest 25%rcSsg|] }|dqS)rrrrrrrrrfcSsg|] }|dqS)rrrrrrrrfz>7.1%z>8.1%zD If actual rises with predicted, the model can pick books to quote.)sklearn.model_selectionr{rUrrrprDvaluesrjsortmaxrmr rJrk)rur\r{rQrRrrrSrrrIZyievZrowsqrxlblZsegZprZacrrrrqs6    (20rq?c stdd\}}}tdd}t|D]\}}||d}|d||d|d7<|dd urp|dnt|d|d|d<|d d ur|d nt|d |d |d <|d |d <q&d d |D}|jdddt|dkrtdt|dd St t||} |d | || d } } dd | D} t t t ddd} | || | g}| D]n\}}}tt| ||dd d df}t tfdd|dD}|||d||d dqntdddt| dt| dtd | d!d!d"d#| d$d!d"d%| d!d!d"d#| d$d!d"td&d | D}td'd | D}td(|d%|d)d*}td+d,d-d.d/d0d1d2d3d4d5d3d6 |d7||d8d9d |Dd:D]&|d;d<fd=d |Dqtd>td?d S)@aTEMPORAL split, scored in dollars. Cross-validation leaks here: legs from the same book, the same day and the same regime land in different folds, and a gate is a decision made BEFORE the market it is judged on. So train on the earliest books and test on the latest, and score the test set the way a bankroll would feel it -- realised P&L of the books the gate would have admitted, against the two baselines that already exist: quote everything, and the live volume gate. Tr|cSsgdddddS)Nr )idxr8r7r9r5rrrrrreszbacktest..r-rr8r7Nr9r5cSs.g|]&\}}t|ddkr|d||fqS)rrr7)r)rrbrrrrrfzbacktest..cSs|dS)NrrrdrrrrerfrgrVrz# full books -- too few to backtest.cSs$g|]\}}}|dD]}|qqS)rr)r_rrxrrrrrfrYr2rZrc3s|]}|dkVqdS)rNr)rrx)rRrr rfzbacktest..)rr8compr9 z@================================================================z TEMPORAL BACKTEST (train z books, test )z train rz%Y-%m-%dz .. raz test cSsg|]\}}}|dqSr5rrrrrrrrrfcSsg|]\}}}|dqSrrrrrrrrfz regimes train c Ss|s td|dddddStdd|Dd}dtd d|Dt|}td|ddt|d d |d d |dd|t|dd dS)Nr`26r]z0 booksz>9css|]}|dVqdS)r8Nrrrrrrrfz(backtest..row..r1css|]}|dVqdS)rNrrrrrrrf>5z books z>5.0fz % complete z>+8.2fz $ z>+7.3fz $/book)rjrHr)rZselr8Zcrrrrrows  zbacktest..rowz Zrulerr]rrz Zcompleter^z zP&Lz per bookzquote everythingzlive gate: vol >= 500cSsg|]}|ddkr|qS)r9irrrrrr%rf)g?g333333?g?g?g333333?zmodel: P(all 3) >= z.2fcsg|]}|dkr|qS)rrr)thrrrr'rfzE The gate can only ever REMOVE books, so it wins by cutting losers.z> Judge it against "quote everything" on the same test window.)rUrrmrDrGrCrrrjrKrrrrnr rJrr}rcollections_counter)Z train_fracrQrrmetarrxmrrvZcutZtraintestZtr_idxruZscoredr7rrrZtr_regZte_regrr)rrRrbacktests^   ** (  $     $rcCs2tt}|D]}||d7<q tt|Srb)rrKdictrorC)Zxsryr rrrr,sr__main__)F)r)__doc__r<r@r?rr collectionsrpathlibrZnumpyrJZsklearn.linear_modelrrrrZsklearn.pipelinerZsklearn.preprocessingrrAr:rrrUrzrqrr__name__rrrrs.      b)' B