o j@sdZddlmZmZddlmZddlmZddlm Z ddl m Z ddl m Z ddlmZdd lmZmZeGd d d Z ddedededededededefddZd S)a prediction_pipeline.py Live wiring for this session's engine chain -- stream_features -> trend_engine -> structure_engine -> regime_engine -> setup_engine -> prediction_engine -> entry_score.py -- the live counterpart to simulate_prediction_pipeline.py's offline replay, same relationship as regime_strategies.py (live-callable) vs simulate_regime_strategies.py (offline harness). evaluate_entry_pipeline() is monitor.py's alternative to entry_engine.evaluate_entry(), selected via config.json's entry.decision_engine == "prediction_pipeline" (default stays "legacy" -- entry_engine.evaluate_entry() is unchanged and still the default path). It returns an EntryDecision-shaped result (same field names as entry_engine.EntryDecision: should_enter, confirmation_score, reasons_for, reasons_against) so monitor.py's call site only needs to choose which function to call, not restructure around a different shape. CALLER-HELD STATE: same contract as every engine built this session -- this module holds nothing itself. monitor.py keeps one dict per symbol (Monitor._pipeline_state[symbol]) across poll cycles and passes it in as `state`; this function mutates it in place (prior_features/ prior_prediction/confirmation_state) and also returns it for convenience. A fresh {} for a symbol (first call ever, or a deliberate reset) is fine -- everything defaults to a cold start, exactly like the offline replay's per-symbol loop. DELIBERATELY LOW EXPECTED FIRE RATE, KNOWN GOING IN: four days of offline replay (2026-09-01 through 09-04, see data/snapshots// prediction_pipeline_backtest.json) found 3 entries across 4 days / 405 symbol-days -- prediction_engine.py's score rarely clears entry_score's min_prediction_score threshold. That backtest ran with bars_sub=[] throughout (the Yahoo snapshot data has no real sub-minute ticks) -- LIVE, stream.py's real sub-minute buffer means velocity_sub/slope_sub/ trade_flow/pressure are actually populated here, which several prediction_engine.py components read directly and default to neutral values without. Live behavior may therefore differ from the backtest in either direction -- exactly why this is a paper-mode test, not a conclusion drawn from the backtest alone. ) dataclassfield)compute_features)classify_trend)analyze_structure)classify_regime)evaluate_setups)compute_prediction)compute_entry_score ACTION_READYc@sJeZdZUeed<eed<eed<eedZ eed<eedZ eed<dS)PipelineDecisionsymbol should_enterconfirmation_score)default_factory reasons_forreasons_againstN) __name__ __module__ __qualname__str__annotations__boolfloatrlistrrrrO/var/www/screener/trade/premarket_backup_2026-09-08_2010/prediction_pipeline.pyr 6s r Nr barsbars_subavg_vol_baselinesession_elapsed_minutesstatesub_bucket_secondsreturnc  Cs|dd|dd|dit||||||d||d} t| } t||} t|| | | |} t| | | |} t|| | | | | |dd}t|| | | || | |d|d }| |d<||d<|j|d<|j t k}t |||j |rp|j ng|rwgdS|j dS) a bars / bars_sub / quote: same shapes stream.StreamManager already hands entry_engine.evaluate_entry() (bars, quote) plus the new get_bars_sub() this session added. avg_vol_baseline: same "expected normal volume" baseline entry_engine/intraday_health already use. session_elapsed_minutes: market_time.minutes_since_open() -- drives regime_engine's OPENING_VOLATILITY window check. state: caller-held per-symbol dict, see module docstring. prior_featuresNprior_predictionconfirmation_state)r priorr#as_of)regimesetupsr()r*r+r'r))r rrrr) setdefaultrrrrrr r confirmationactionr r entry_scorereasons)r rrquoter r!r"r#r)featurestrend structurer*r+ prediction entry_readingrrrrevaluate_entry_pipeline?s:         r7)rN)__doc__ dataclassesrrstream_featuresr trend_enginerstructure_enginer regime_enginer setup_enginerprediction_enginer r/r r r rrrdictintr7rrrrs.*