o Vj@stdZddlmZmZdZdZddddd d ZeGd d d Zd dZ dde de de de de de defddZ dS)a exit_velocity_experimental.py [2026-09-19] NOT wired into monitor.py's live loop -- backtest-only. Built from the cross-symbol peak-pattern analysis run against the full 2026-09-18 session (all 30 candidates, not just the 9 the bot actually traded): neither imbalance level at a stalled peak nor imbalance flipping negative afterward showed any real discriminating power between the 161 failed-breakout events and the 275 that recovered (52-73% recovery, non-monotonic, across every imbalance bucket tested) -- directly undermining the assumption exit_prf_experimental.py and exit_resistance_timer_experimental.py were both built on. The one signal that DID show a clean, monotonic, statistically significant gradient (ratio>=1.0 vs <1.0: 73.0% recovery n=204 vs 54.3% n=232, p~0.00005) was the ratio of trade participation (buy+sell volume) in the 90s after a stall vs the 90s before it -- rising participation predicts recovery, fading participation predicts failure. This does NOT explain GLOO itself (participation ratio 0.92, imbalance +73% at its own real top -- looked completely healthy right up to the reversal), so it is not being sold as "the fix." It's a different SHAPE of signal than everything tried so far, per the explicit design goal this was built to: PRF and the resistance-timer both wait a FIXED time window (poll-count confirm reads, or a flat 60-120s clock) regardless of how fast or slow price is actually giving back. This module instead reacts to the RATE of the post-peak decline (a short trailing-window price velocity, independent of how long the stall has been running) -- matching the one clear practitioner rule the accompanying web research turned up for this exact discrimination problem: a fast snapback is a failure, a move that holds/builds is a genuine pause -- gated by the participation-ratio finding above so a merely-noisy fast wiggle with volume already picking back up doesn't fire. Single-read trigger once both conditions are true (no confirm-read hysteresis layered on top, deliberately -- the whole point is reacting faster than PRF/resistance-timer's wait-N-reads designs), using stream.py's 30s sub-bars (get_bars_sub()) for both the velocity reference price and the participation-ratio volumes, same buffer fast_slope_exit.py already reads. ) dataclassfieldHOLDEXITg@g@g?) min_mfe_pctvelocity_lookback_buckets min_decline_velocity_pct_per_minparticipation_window_bucketsmax_participation_ratio_to_exitc@s^eZdZUeed<eZeed<dZeed<dZ eed<e e dZ e ed<e e dZ e ed <d S) VelocityDecisionsymbolstateF should_exitreason)default_factorymetrics state_outN)__name__ __module__ __qualname__str__annotations__ STATE_HOLDrrboolrrdictrrrr5/var/www/screener/trade/exit_velocity_experimental.pyr 8s    r cCsit|piSN)DEFAULT_CONFIG)cfgrrr _merge_cfgBsr#Nrbarssub_bars entry_pricestate_inr"returncCst|}t|pi}t|d|ddd}t|dks#t||kr*t|d|dS|dd }|d |}t||} ||k} |rI| ||d nd } | |d k} d | i} |t| dt| d| | d}| sxt|td| dd|d d|| dS| rt|td|dd|| dS|d|d}|d }|dd}|r|||d nd }|r||dnd }|d}tdd|| dD}t|d|krtdd|d|| Dnd }|dkr||nd}| t|dt|dt|d|durt|dndd ||d!k}|duo||d"k}|r4|r4t|t d#d$| dd%|d&d'|d(d)|dd* || d+St|td,| dd-|d&d.|d!d/|durL|nd0d1|d"d || dS)2a bars: same short rolling 1-min-bar window exit.py/PRF take. sub_bars: stream.SymbolBuffer.get_bars_sub() -- 30s-bucket bars, oldest first, each {"t","o","h","l","c","v",...}. state_in: caller-held state fed back every poll -- {"peak_price"}. Fresh {} at entry. No confirm-count/stage tracking needed -- this is a memoryless single-read trigger past the peak itself. r r rrzinsufficient sub-bar history)rrrc peak_pricegY@gr)pricer,mfe_pctarmed made_new_highznot armed yet (MFE z.2fz% < z%))rrrrrzmaking new highs ($z.4f)g>@gN@cs|]}|dVqdSvNr.0brrr {zevaluate..Ncsr3r4rr6rrrr9|r:r)velocity_pct_per_min recent_vol prior_volparticipation_ratior r Tzfast decline off peak $z: z.1fz %/min over z.0fzs, participation ratio z (not picking back up))rrrrrrz off peak $z : velocity z%/min (need >=z), participation ratio zn/az (need <) r#rmaxlenr getroundrsumupdate STATE_EXIT)rr$r%r&r'r" need_bucketsr. prior_peakr,r1r/r0 base_staterref_bar ref_pricelookback_seconds velocity_pctr<wr=r>r?decline_fast_enoughparticipation_not_elevatedrrrevaluateFs      4    rQr )__doc__ dataclassesrrrrFr!r r#rlistfloatrrQrrrrs*)