ށj)DdZddlmZddlZddlZddlZddlmZmZddl m Z ddl Z ddl Ze ej!j"j"Zej&j)deeddlmZej0dZej0dZej0d Zd gej;eZed Gd dZddZ ddd dZ!d!dZ"d"dZ#d#dZ$d$dZ%d%dZ&d&dZ'dddddd'dZ(d(dZ)e*dk(r e+e)y))aCausal, non-promoting study of MLP feature states before extreme BTC bars. The event at bar t is the close-to-close return from t-1 to t. Every candidate predictor is taken from t-1, so it was available before the event bar began. Tail thresholds and feature cuts are calibrated only in the discovery window, then held fixed in validation and OOS windows. This is an event-profile diagnostic, not a training command and not a source of promoted MLP parameters. ) annotationsN)asdict dataclass)Path) FEATURE_COLSz2023-12-31 23:59:59z 2024-01-01z 2026-03-01 mlp_scoreT)frozenc,eZdZUded<ded<ded<y)TailDefinitionfloat up_return down_return tail_fractionN)__name__ __module__ __qualname____annotations__(tools/run_mlp_extreme_bar_event_study.pyr r !srr c0tj|d}|jjr1tj|ddj j dStj|ddj j dS) zLParse TradingView epoch-second exports and ISO timestamps deterministically.coerce)errorssT)unitutcNraise)rr)pd to_numericnotnaall to_datetimedt tz_localize)valuesnumerics rparse_chart_timer'(slmmF84G}}~~gCT:==II$OO >>&d7 ; > > J J4 PPr@ bar_hoursc|dkr tdddg|Dcgc]}||jvs|}}|rtd||ddg|j}t|d|d<|j dj d}|dj js|djs td|djd }tj|d z }||dz |k}tj|djd |dz j||d <|j||d <|Scc}w)zCReturn feature snapshots at t and their contiguous next-bar return.rzbar_hours must be positivetimeclosezmissing required columns: T)dropz+time must be unique and strictly increasing?)hoursforward_return outcome_time) ValueErrorcolumnscopyr' sort_values reset_index duplicatedanyis_monotonic_increasingshiftr Timedeltanplogwhere) datafeaturesr*columnmissingframe next_timemax_gap contiguouss rcausal_event_framerI0sfA~566%+W$@x$@_$@&FRVR^R^D^v$@G_5gY?@@ &'-H- . 3 3 5E$U6]3E&M   f % 1 1t 1 '**=9:# rctjd|j}| ||d|k\z}|||d|kz}||d|kz}|j||dj zj S)NT)indexr,r3r2)rSeriesrTlocr r6)rEstartendmasks r_windowrZRs 99T -D  f &&  f $$ n%,, 99TE"2399;; < A A CCrc|dk(r|d|jk\n|d|jk}|dk(r|||k\n|||k}||j}||z}|jrt ||j nd} |jrt ||j nd} t |jt |j| | | | r | dkDr| | z dSddS)Nupr2highr) observationsselected event_rate base_ratelift)r rr r:r meanintsum) rEfeaturecutsideeventtails event_maskr_usablerar`s r_profile_windowrm\s?D}'(EOO;RWXhRimrm~m~R~J(,uW~$E'Ncr\downzevent must be 'up' or 'down'rKz# has too few discovery observationsrLlowr]rb-inf)rfrirhrgrP validationoos)r4rMrNr rOrm)rPrrrsrfrirj feature_tailr%low_cuthigh_cutrpr]low_lift high_liftrhrgdiscovery_profiles rprofile_featurerzls# N"788 w  & & (F 6{RG9$GHIIfool;hXG )WgueU KC 9gx NDf+6vV 8Uf iH?HH?TFHd#;[`biknZo D# &%j'3eUSsGS$uE rc|d}t|tsJt|dxs tdt|dfS)NrPrbrqrf) isinstancedictr str)profilerPs r _rank_keyrsD $I i && & )F#4uV} 5s79;M7N OOrct||}|||d}i}dD]} g} |D]} t|||| | ||} |jD cic]8\} }| t|| t | dt | d| t||:}} }| j | | | d| d|dt| dd || <|Scc}} w) a6Describe within-window tail behavior without reselecting feature rules. The event thresholds are recalculated within each reporting window, so this is descriptive of *relative* extremes rather than an executable fixed-size return rule. Feature direction and cut remain selected in discovery. rPrrrsr\rorfrirjrtrgrh)rfrirhrgwindowscDt|dddxs tdS)NrrPrbrq)r )rs rz(relative_tail_profiles..s$gi&8&Ef&M&^QVW]Q^ _rTkeyreverse)rRrzitemsrmr r~appendsorted)rPrrrsrBrtrdiscovery_tailsrresultriprofilesrfr_namewindowrelatives rrelative_tail_profilesrs &i?O%ZLG13FG&:sG5%LH%,MMO%4LD&ofguXe_7MsS[\bScOdfk&5fm&LNN$3  OO" (#   " _ u ' 0 M#s=B; 皙?皙?rrtr*rBc tj|}|jjj jj |_t tj|xs#tDcgc]}||jvs|c}}|s tdt|||}t|dt} t|tttj dz } t|td} t#| |} i} dD]6}|Dcgc]}t%| | | ||| |}}t'|t(d| |<8t|d |d zt+| |tttttttj dz gttd t-| t-| t-| d | t/| | | ||| dScc}wcc}w)Nz(no requested feature columns are presentr)) nanosecondsrrTrzFnext 6h close-to-close return; feature snapshot is the preceding closer0)rimax_label_gap_hourstail_definitionrt discovery_endrr oos_startr)rBrtr) data_path definitionrstudiesrelative_tail_studies)rread_csvr5r~lowerstriplistr}fromkeysDEFAULT_FEATURESr4rIrZ DISCOVERY_ENDVALIDATION_START OOS_STARTr=rRrzrrrrNr)rrrtr*rBrArfselected_featuresrErPrrrsrjrrirs rrunrs ;;y !D<<##))+//557DLT]]8+{O_7{O_GcjnrnznzczO_7{|} CDD t%6) LEt]3I 0)bllWX>Y2YZJ %D )C I} 5E24G#46#4$Iz3W\diyEF#4 6iF ^]#,s?%e}( //0#i",,[\B]6]2^_Y "%Ys:WZ[^W_`!7 z31B%]"  8|6s7G, G, G1ctjt}|jdtdd|jdt d|jd t d |jd t d d |jddd|jdtd|j }t|j|j|j|j|j}tj|d}t||j rE|j j"j%dd|j j'|dzy)N) descriptionz--dataTz4TradingView chart-data CSV to read without modifying)typerequiredhelpz--tail-fractionr)rdefaultz--feature-tailrz --bar-hoursr(zBExpected chart-bar duration; labels crossing >1.5 bars are dropped)rrrz --features*z@Feature columns to profile; defaults to mlp_score when available)nargsrz--outputzOptional JSON output path)rrr)indent)parentsexist_ok r)argparseArgumentParser__doc__ add_argumentrr parse_argsrrArrtr*rBjsondumpsprintoutputparentmkdir write_text)parserargsrtexts rmainrsB  $ $ 9F tdAwx )tD (udC  E3FJK  C6xy  4OP    D $*<*<4K\K\hlhvhvBFBOBOPF ::fQ 'D $K {{    = td{+ r__main__)r% pd.Seriesreturnr)rA pd.DataFramerB list[str]r*r rr)rPrrr rr )rErrWpd.Timestamp | NonerXrrr)rErrfr~rgr rhr~rir~rjr rzdict[str, float | int | None])rPrrrrrsrrfr~rir~rjr rtr rdict[str, object])rrrztuple[float, str])rPrrrrrsrrBrrtr rr rz"dict[str, list[dict[str, object]]]) rrrr rtr r*r rBzlist[str] | Nonerr)rrd),r __future__rrrsys dataclassesrrpathlibrnumpyr>pandasr__file__resolverREPOpathinsertr~strategies.strategy_mlp_scoresr Timestamprrrr}rrr r'rIrRrZrmrzrrrrr SystemExitrrrrs/ # ) H~&&--3t97 23 2<< - BLL & >$-- "=> $ QWZ( D  .P "J48thkKO!H $ z TV r