Warrior_EA/research/vplevels.py

"""Turn raw (day, price-bin, weight) profiles into the operational levels the books use.

Every level here is STRICTLY CAUSAL: a level in force on broker-day d is computed only from
days < d. That is not a formality. A session VPOC is the single most tempting lookahead in
this whole subject, because the natural way to draw it on a chart is over the day you are
looking at - and a level fitted to the day it is tested on will "hold" beautifully.

Levels produced
---------------
  vpoc / vah / val   previous session's point of control and value area edges
  naked vpoc         a past session VPOC that price has NOT traded through since. Held to
                     be an unfinished auction and therefore a magnet.
  hvn / lvn          peaks and valleys of a rolling composite profile over the previous
                     N sessions. HVN = agreement, claimed to attract and to make good
                     targets; LVN = rejection, claimed to make good stop locations.
"""
import numpy as np
from scipy.signal import find_peaks

BROKER_OFFSET_MS = 2 * 3600 * 1000
DAY_MS = 86400 * 1000


def broker_day(t_ms):
    return ((np.asarray(t_ms, np.int64) + BROKER_OFFSET_MS) // DAY_MS).astype(np.int64)


def load(sym, path='c:/Users/admin/Documents/Workspaces/Market Data/profiles/'):
    z = np.load(f"{path}{sym}_vp.npz")
    return z['day'], z['bin'], z['ticks'], z['dwell'], float(z['binsize'][0])


def day_index(day, b, w):
    """Group the sparse cells by day. Returns (days, list of (bins, weights)) sorted."""
    order = np.lexsort((b, day))
    day, b, w = day[order], b[order], w[order]
    starts = np.concatenate(([0], np.flatnonzero(np.diff(day)) + 1))
    ends = np.concatenate((starts[1:], [len(day)]))
    return day[starts], [(b[s:e], w[s:e]) for s, e in zip(starts, ends)]


def densify(bins, w):
    """Sparse bins -> contiguous array with explicit zeros for untraded levels inside the
    range. Without the zeros a value-area walk annexes across a hole as if it were adjacent
    and low-volume nodes cease to exist - which is the whole point of them."""
    full = np.arange(bins[0], bins[-1] + 1)
    dense = np.zeros(len(full))
    dense[bins - bins[0]] = w
    return full, dense


def session_levels(days, cells, frac=0.70):
    """Per-session VPOC / VAL / VAH / high / low, all as integer bin indices."""
    from profiles import value_area
    out = np.empty((len(days), 6), np.int64)
    for i, (bins, w) in enumerate(cells):
        full, dense = densify(bins, w)
        poc, lo, hi = value_area(full, dense, frac)
        out[i] = (days[i], poc, lo, hi, bins[0], bins[-1])
    return out          # day, vpoc, val, vah, low, high


def composite_nodes(days, cells, lookback=20, smooth=5, prominence=0.35):
    """HVN/LVN from a rolling composite of the PREVIOUS `lookback` sessions.

    Smoothing first is not cosmetic: on a raw 0.1-pip grid every second bin is a local
    maximum, and "find the peaks" without it returns a few hundred meaningless nodes per
    day, which would let any subsequent test claim a hit against whatever price did.
    `prominence` is in units of the smoothed profile's own standard deviation, so it means
    the same thing on gold and on EURUSD.

    Returns {day: (hvn_bins, lvn_bins)} for every day that has a full lookback behind it.
    """
    out = {}
    for i in range(lookback, len(days)):
        bb = np.concatenate([cells[j][0] for j in range(i - lookback, i)])
        ww = np.concatenate([cells[j][1] for j in range(i - lookback, i)])
        k, inv = np.unique(bb, return_inverse=True)
        agg = np.bincount(inv, weights=ww, minlength=len(k))
        full, dense = densify(k, agg)
        if len(full) < 4 * smooth:
            continue
        ker = np.ones(smooth) / smooth
        sm = np.convolve(dense, ker, mode='same')
        p = prominence * sm.std()
        hv, _ = find_peaks(sm, prominence=p)
        lv, _ = find_peaks(-sm, prominence=p)
        out[int(days[i])] = (full[hv], full[lv])
    return out


def naked_vpocs(sess, max_age=60):
    """For each day, the past VPOCs price has not traded through yet.

    A VPOC is 'naked' until a LATER session's range covers it. Only sessions strictly
    before the current one can retire it, so the answer for day d never depends on day d.
    Capped at `max_age` sessions back: a naked VPOC from three years ago is not an
    unfinished auction, it is a price the market has left behind, and including it would
    load the test with levels no trader would draw.

    Returns {day: array of still-naked vpoc bins}, nearest-in-time first.
    """
    out = {}
    vpoc = sess[:, 1]; lo = sess[:, 4]; hi = sess[:, 5]
    for i in range(1, len(sess)):
        j0 = max(0, i - max_age)
        cand = vpoc[j0:i]
        #--- covered by any session strictly between its own and today
        alive = np.ones(len(cand), bool)
        for k in range(len(cand)):
            src = j0 + k
            if src + 1 >= i:
                continue
            rl = lo[src + 1:i]; rh = hi[src + 1:i]
            if np.any((rl <= cand[k]) & (rh >= cand[k])):
                alive[k] = False
        out[int(sess[i, 0])] = cand[alive][::-1]
    return out