js dZddlmZddlmZddlZddlZddl m Z ddl m Z ddl mZddlmZdd lmZed Gd d ZddZddZdddddZdddd ddZddZy)zILeakage-aware research utilities for relative extreme-bar classification.) annotations) dataclassN) SimpleImputer)LogisticRegression)average_precision_score) make_pipeline)StandardScalerT)frozenc,eZdZUded<ded<ded<y)TailRulestrfeaturesidefloatcutN)__name__ __module__ __qualname____annotations__%strategies/mlp_tail_event_research.pyr r s L I Jrr cFd|cxkrdkstdtd|d}|j||jd|z }}tjd|jt j }d|j||k<d|j||k\<|S) zCLabel each complete outcome as down (-1), ordinary (0), or up (+1).r?z'tail_fraction must be between 0 and 0.5forward_return?)indexdtype) ValueErrorquantilepdSeriesrnpint8loc)frame tail_fractionreturnslowerupperlabelss rrelative_tail_labelsr.s } "s "BCC #BCC$%G##M2G4D4DS=EX4Y5E YYq 277 ;F#%FJJw% #$FJJw% Mrc|jdk(r||j|jkS|jdk(r||j|jk\Std)Nlowhighz!rule side must be 'low' or 'high')rrrr!)r(rules r _selectionr3"sV yyET\\"dhh.. yyFT\\"dhh.. 8 99r rvol_norm )volatility_featurebinsc @|dkr tddd|j|g}|Dcgc] }||vs| }}|rtd||j|j}t ||} t |||d<| j d|d <|djj|d <|jd |jd d j|j|dz jt|d<dx} x} x} } d}|jd dgd D]\}}||d||d}}|js |jr0t!|}| |z } | t!|z } | ||d j#zz } | ||d j#zz } |dz }| dk(r ddddddddS| | z }| | z }t| t| |t%|t%|t%||z |dkDrt%||z dSddScc}w)u Compare selected and control bars within calendar-year × volatility bins. This standardizes the control rate to the selected bars' stratum mix. It is deliberately a simple matched descriptive test, not a causal estimator of a deployable trading rule. zbins must be at least twotimermissing required columns: subsetselectedrdownyearfirstT)methodpctg-q=)r,vol_bingr)observedr N)r>controlsstrata selected_ratematched_control_ratedeltalift)r!rdropnacopyr.r3eqdtr@groupbyrankmulclipastypeintemptylenmeanr)r(r2r)r6r7requiredcolumnmissingdatar-selected_total control_totalweighted_selected_rateweighted_control_raterG_groupr>controlweightrH control_rates r#volatility_matched_downside_profilerf*s ax455($,,8JKH$,DH&e0CvHGD5gY?@@ <CC7X\C]aabfgllswzslAHHILMDOVYYNY]Y%;>S FLL&)!4tLD5!% "34eU:=N>W]] X& W% &8F+;+@+@+B"BB'&/*>*>*@!@@! E1Djny}HLM M*^;M(>9L' &}- %l 3}|347Ca7Gml23 NR 5Es HH) block_bars repetitionsseedc p|dkr td|dkr td|jdd|jdgjd }t ||kr td t j j|}g}t jt ||z d z} t|D]} g} t | t |krPt|j| } | jt| | |zt | t |krP|j| d t |j} t| ||d}|t j |s|j#t%||s |dd d d d dSt j&|}|t |t%t j(|dt%t j(|dt%t j(|dt%|dkDj+dS)azEstimate matched-lift uncertainty with deterministic contiguous resampling. Resampling individual bars would assume independent six-hour returns. This uses contiguous blocks (seven days by default), retaining short-run serial structure. It is an uncertainty description for this observational study, not an optimization input or a trading confidence score. r9zblock_bars must be at least twozrepetitions must be at least 20r:rr4r<T)dropz'frame needs at least one complete blockr N)r)rKr)rkvalid_repetitionsp05p50p95share_above_oneg?rgffffff?r)r!rLr reset_indexrWr%random default_rngarangerangerUchoiceextendilocrMrfisfiniteappendrasarrayr"rX)r(r2r)rjrkrlr\rngliftsstartsraindexesstartsamplerKvaluess r#moving_block_bootstrap_matched_liftrYs A~:;;R:;; <<(8$,, T< U a agk a lD 4y:BCC ))   %CE YYs4y:-1 2F ; 'lSY& 6*+E NN5 (:; <'lSY&7:CI./446264}]^de   D 1 LLt %  *4X\eiCD D ZZ F" [R[[./R[[./R[[./ &3,!4!4!67  rc g|dDcgc] }||vs||vs |}}|rtd|t||jdz}t||jdz}tt dt t dddd d }||jtjtj gtj} ||jtjtj gtj} |j| ||d } |d d j| } tjd| | j| jz} | | j!ddz} tj"| }||j%ddz }| j&}t)tj*|dk(dt)tj*|dk(d}}t-|t/t1|dk(|d d |ft/t1|dk(|d d |ft/|dScc}w)zDFit on an earlier window and score relative tails in a later window.rr;r median)strategyg?balancedirisaga)C class_weightmax_iter random_statesolverrNz ij,kj->ikT)axiskeepdimsrr9) observations down_pr_auc up_pr_auctail_base_rate)r!r.to_numpyrrr rreplacer%infnanfit transformeinsumcoef_ intercept_maxexpsumclasses_rU flatnonzerorWrr)train evaluationfeaturesr)rZr[y_trainy_evalmodeltrain_featuresevaluation_features classifier transformedscores exp_scores probabilitiesclasses down_column up_columns rtail_classifier_metricsrs7$Ah$A0@$Au$A&VSXEX\bjt\tv$AGu5gY?@@"5-8AACaGG !*m < E E G! KF x( #   E"8_,,bffrvvg->GN$X.667H"&&Q IIng& rJ*&&':;K YY{K1A1A BZEZEZ ZF fjja$j//FJQ!FFM!!G 1 !=a!@A3r~~V]abVbGcdeGfCgKJ4Vq[-PQS^P^B_`a26Q; aQZl@[\] .  Kvs II)r( pd.DataFramer)rreturn pd.Series)r(rr2r rr) r(rr2r r)rr6r r7rUrdict[str, float | int | None])r(rr2r r)rrjrUrkrUrlrUrr) rrrrrz list[str]r)rrzdict[str, float | int])__doc__ __future__r dataclassesrnumpyr%pandasr#sklearn.imputersklearn.linear_modelrsklearn.metricsrsklearn.pipelinersklearn.preprocessingr r r.r3rfrrrrrrsO"!(33*0 $ :ALZ\,h- - - -  -  - -#-`,r