\jJSrSSKrSSKrSSKrSSKrSSKrSSKrSr Sr SSSS.r S\RS \RS \ 4S jrS \4S jr\S :Xa\R$"\"55 gg)u Scoring CLV dell'agente formazioni (e futuri agenti live). Per ogni segnale in agent_signals con partita giocata: - snapshot: le probabilità di mercato più recenti PRIMA del segnale (odds_snapshots; fallback: quote pre-match in matches) - closing: le probabilità implicite di chiusura (imp_prob_* in matches) - direzione: sull'esito dove l'agente ha spostato di più la probabilità, la closing si è mossa nella stessa direzione dello spostamento? KPI primario (dal design): CLV direction hit rate. >55% = l'agente vede cose vere prima del mercato; <=52% dopo 150 segnali non nulli = kill criterion. KPI secondario: log loss delle probabilità aggiustate vs snapshot vs closing. Uso: python scripts/score_signals.py Ng{Gzt?gMb`?)HDApoutcomereturnc [R"USS5nXRSSS9- n[[R"[R "U[R "[U55U455*5$)Ng-q=rT)axiskeepdims)npclipsumfloatmeanlogarangelen)rr s 0C:\wamp64\www\Tigertips\scripts\score_signals.pyloglossr!s` 5!A EEq4E ((A "''"&&299S\#:G#C!DEFF GGc x[R"SS9n[R"U5 UR 5n[ R "URURURURSS9n[R"SU5nUR5 UR(a [S5 gUR R#/SQ5Hn[R$"X4S S 9X4'M US R'5US R'5-[(:US '[S[+U5SUS R-5S35 UR/SS/5GHuupVn[SUSUS[+U5SUS R-5S3 5 XwS R1/SQS9R35n[+U5(Ga}U/SQR55n U/SQR55n U/SQR55n X- n [6R&"U 5R9SS9n [6R:"[+U55n[6R<"XU 45nXU 4XU 4- n[6R&"U5[>:n[6R<"UU)5UU):HnUR@(aRURC5nUS:aS O US!:aS"OS#n[S$US%S&UR@S'UR-5S(U35 UR@S):aURC5S*::a [S+5 O [S,5 UR1/S-QS9R35nUUS.RE[F5n[+U5S/:aUS.RI[F5R55n[S0[+U5S1[KU/SQR55U5S2S3[KU/SQR55U5S235 [5 GM g)4NzScoring CLV dei segnali agente) description tigertips)hostportuserpassworddatabaseaG SELECT s.id AS signal_id, s.agent, s.stage, s.checked_at, s.delta_lambda_home, s.delta_lambda_away, m.result, m.imp_prob_home AS close_h, m.imp_prob_draw AS close_d, m.imp_prob_away AS close_a, pm.prob_home AS model_h, pm.prob_draw AS model_d, pm.prob_away AS model_a, pa.prob_home AS adj_h, pa.prob_draw AS adj_d, pa.prob_away AS adj_a, os.prob_home AS snap_h, os.prob_draw AS snap_d, os.prob_away AS snap_a FROM agent_signals s JOIN matches m ON m.id = s.match_id AND m.result IS NOT NULL LEFT JOIN predictions pm ON pm.match_id = s.match_id AND pm.stage = s.stage AND pm.source = 'model' LEFT JOIN predictions pa ON pa.match_id = s.match_id AND pa.stage = s.stage AND pa.source = 'agent_adj' LEFT JOIN odds_snapshots os ON os.id = ( SELECT os2.id FROM odds_snapshots os2 WHERE os2.match_id = s.match_id AND os2.taken_at <= s.checked_at ORDER BY os2.taken_at DESC LIMIT 1) zENessun segnale con partita giocata: lo scoring parte con la stagione.r)agentstage checked_atresultcoerce)errorsdelta_lambda_homedelta_lambda_awaynonnullzSegnali con esito: z | non nulli:  r!r"z=== z / z (z segnali, z non nulli) ===)snap_hadj_hclose_h)subset)r+snap_dsnap_a)r,adj_dadj_a)r-close_dclose_ar)r g?z&POSITIVO: l'agente anticipa il mercatogQ?zin linea col rumoreNEGATIVOz CLV direzionale: z.1%z su z segnali (piatti esclusi: z) -> gp= ף?z: *** KILL CRITERION RAGGIUNTO: valutare archiviazione ***z> CLV direzionale: dati insufficienti (serve snapshot+closing))r,r-r$r$ z log loss (z partite): aggiustate z.4fz | closing )&argparseArgumentParserconfig add_db_args parse_argspymysqlconnectrrrrpdread_sqlcloseemptyprintcolumnsdrop to_numericabs MIN_DELTArrgroupbydropnacopyto_numpyrargmaxrsign FLAT_MOVEsizerisin RESULT_INDEXmapr)parserargsconndfcr!r"grpgsnapadjrAshiftcompidx agent_dir close_moveflathitsrateverdictg2ys rmainrh's  $ $1Q RF v    D ??   $(MMK ID & ' B( JJL xx UV ZZ__G H beH5I+,002+,00236?@ByM Byr)}7H7H7J6K2 NO!zz7G*<= UG3ugRCzY##%&o7 8 I  & &1 ' 33746 q6634==?D/099;C78AACEJE66%=''Q'/D))CF#C4i 01IDy)DdO;J66*% 1D77:te,-D51AADyyyy{GKd{C9= 5 *+D:T$))E**.((*U7)EFyyCDIIK4$7RS R SZZ=Z > C C E 8 !!,/ 0 r7b=8   .779A LR * '+F(G(P(P(RTU VWZ[\$R(I%J%S%S%UWXYZ]^` a M >N r__main__)__doc__r8sysnumpyrpandasr?r=r:rHrOrRndarrayrrintrh__name__exitrrrss$    QQ' HrzzHBJJH5H Q cQ h zHHTVr