Warrior_EA/research/sqx.py

"""Decoder for StrategyQuant X tick .dat files (format 4.2).

Derived from SQX's own writer, disassembled out of internal/libs/SQDataLib.jar:
  com.strategyquant.datalib.data.io.newDataFormat.TickDataWriter
  com.strategyquant.datalib.data.io.newDataFormat.NewDataFormat{,Writter}

Why bother, when Scripts/ExportTicks.mq5 pulls the same four fields from MT5: DEPTH. The
broker's MT5 tick history goes back a few years; these files start in 2011. Sample size has
been the binding constraint on this entire investigation, so 15 years of ticks is worth a
decoder.

FORMAT
------
Header: four Java writeUTF strings (version "4.2", a type char, a column string, and an
empty one), ten zero bytes, then one more writeUTF. Data follows.

Records are (time, ASK, BID, volume) - ask before bid, and the writer swaps them if bid>ask,
so ask is always the larger. Every BLOCK_LENGTH=1000 records a block header is emitted:

    MAGIC (15 bytes: 0x00..0x0e) + int32 block index + config + four raw int64 values

and in between, delta records:

    config + four variable-width deltas against the previous record

CONFIG is two bytes = four nibbles, one per field, laid out high-nibble-first:

    nibble = (logicType << 2) | dataType
    dataType : 0=BYTE(1) 1=SHORT(2) 2=INT(4) 3=LONG(8)      <- chosen by |value| magnitude
    logicType: 0=MINUS   1=PLUS     2=ASIS                   <- sign carried here, not in the bytes

MINUS/PLUS payloads are UNSIGNED magnitudes with the sign supplied by logicType; ASIS is a
plain signed read (used for the full records at block starts).

SCALING
-------
Prices are integers scaled by 10**decimals. SQX carries `decimals` in its own metadata, not
in this file, so it is inferred here by checking which power of ten yields a sane price -
and validated against the first record, which is uncompressed and therefore unambiguous.
Volume divides by VOLUME_CONSTANT = 100000.0 (OLD_VOLUME_CONSTANT = 100.0 for old files).
A volume of Long.MAX_VALUE is the writer's "no volume" sentinel and decodes to NaN.
"""
import struct, sys, datetime as dt
import numpy as np

MAGIC = bytes(range(15))
BLOCK_LENGTH = 1000
VOLUME_CONSTANT = 100000.0
_SIZE = (1, 2, 4, 8)
MINUS, PLUS, ASIS = 0, 1, 2
NO_VOLUME = (1 << 63) - 1


def read_header(b):
    """Returns the offset at which the first block MAGIC starts, plus the header strings."""
    p, strs = 0, []
    for _ in range(4):
        n = struct.unpack_from('>H', b, p)[0]
        strs.append(b[p + 2:p + 2 + n].decode('latin1'))
        p += 2 + n
    p += 10                                  # ten zero bytes
    n = struct.unpack_from('>H', b, p)[0]
    strs.append(b[p + 2:p + 2 + n].decode('latin1'))
    p += 2 + n
    if b[p:p + 15] != MAGIC:                 # be forgiving: just find it
        p = b.find(MAGIC)
        if p < 0:
            raise ValueError("no block magic found in first chunk")
    return p, strs


def _fields(cfg):
    """Unpack the two config bytes into four (logicType, dataType) pairs."""
    out = []
    for i in range(4):
        nib = (cfg[i >> 1] >> (4 if i % 2 == 0 else 0)) & 0xF
        out.append((nib >> 2, nib & 3))
    return out


def decode(path, max_records=None, progress=None):
    """Whole-file decode into four int64 arrays. Convenient, but a full symbol is ~458M
    ticks = ~15 GB of arrays - use decode_iter() for anything beyond a sample."""
    parts = [[], [], [], []]
    for ts, a, b, v in decode_iter(path, max_records=max_records, progress=progress):
        parts[0].append(ts); parts[1].append(a); parts[2].append(b); parts[3].append(v)
    if not parts[0]:
        return tuple(np.empty(0, dtype=np.int64) for _ in range(4))
    return tuple(np.concatenate(p) for p in parts)


def find_block(f, approx, size, window=1 << 20):
    """Byte offset of the first VALID block header at or after `approx`, or None.

    This is what makes the format parallel-decodable: a block header restates all four
    fields as absolute int64s, so a decoder can start there cold with no history. The catch
    is that MAGIC is 0x00..0x0e, a byte run that delta payloads produce by coincidence
    fairly often - a small negative delta on a LONG field is a lot of zero bytes. So a
    candidate is only accepted when the NEXT candidate downstream carries the next
    sequential block index, which junk essentially never does.
    """
    def candidate(off):
        """(offset, block_index, timestamp) of the first plausible header at/after off."""
        p = off
        while p < size:
            f.seek(p)
            buf = f.read(window + 64)
            if len(buf) < 21 + 32:
                return None
            k = 0
            while True:
                k = buf.find(MAGIC, k)
                if k < 0 or k + 15 + 4 + 2 + 32 > len(buf):
                    break
                idx = struct.unpack_from('>i', buf, k + 15)[0]
                t, a, b, _v = struct.unpack_from('>4q', buf, k + 15 + 4 + 2)
                #--- a real header: sane block index, epoch-ms in [2000, 2035], ask >= bid > 0
                if (0 <= idx < (1 << 28) and 946684800000 <= t <= 2051222400000
                        and a >= b > 0):
                    return (p + k, idx, t)
                k += 1
            p += window
        return None

    first = candidate(approx)
    while first is not None:
        nxt = candidate(first[0] + 15 + 4 + 2 + 32)
        if nxt is None:
            return first[0]                       # last block in the file
        if nxt[1] == first[1] + 1 and nxt[2] >= first[2]:
            return first[0]
        first = nxt                               # the first was noise; try the next
    return None


def split_offsets(path, nchunks):
    """Byte ranges [(start, stop), ...] on block boundaries covering the whole file.

    Half-open and contiguous: chunk k's stop IS chunk k+1's start, so every record is
    decoded exactly once with no duplicates and no gaps.
    """
    import os
    size = os.path.getsize(path)
    with open(path, 'rb') as f:
        head = f.read(1 << 16)
        first, _ = read_header(head)
        if nchunks <= 1:
            return [(first, size)]
        cuts = [first]
        span = (size - first) // nchunks
        for k in range(1, nchunks):
            off = find_block(f, first + k * span, size)
            #--- keep it only if it is strictly after the previous cut
            if off is not None and off > cuts[-1]:
                cuts.append(off)
        cuts.append(size)
    return [(cuts[i], cuts[i + 1]) for i in range(len(cuts) - 1)]


def decode_iter(path, batch=2000000, max_records=None, progress=None, start=None, stop=None):
    """Streaming decode. Yields (time_ms, ask_i, bid_i, vol_i) int64 arrays in batches.

    A full symbol here is hundreds of millions of ticks, so nothing that wants the whole
    file can afford to materialise it. Everything downstream aggregates to bars in one pass.

    start/stop: decode only the byte range [start, stop), which MUST begin on a block
    boundary - see split_offsets(). Used to decode one file across many processes, since
    this per-record loop is pure Python and saturates exactly one core.
    """
    f = open(path, 'rb')
    if start is None:
        head = f.read(1 << 16)
        start, strs = read_header(head)
    f.seek(start)

    ts, asks, bids, vols = [], [], [], []
    t = a = bd = v = 0
    buf = b''
    pos = 0
    base = start                          # absolute file offset of buf[0]
    CH = 1 << 22
    eof = False
    n = 0
    while True:
        if len(buf) - pos < 64 and not eof:
            buf = buf[pos:]
            base += pos
            pos = 0
            more = f.read(CH)
            if not more:
                eof = True
            buf += more
        if len(buf) - pos < 2:
            break
        if stop is not None and base + pos >= stop:
            break
        if buf[pos:pos + 15] == MAGIC:
            pos += 15 + 4                     # magic + int32 block index
            if len(buf) - pos < 2 + 32:
                break
            cfg = buf[pos:pos + 2]; pos += 2
            t, a, bd, v = struct.unpack_from('>4q', buf, pos)
            pos += 32
        else:
            cfg = buf[pos:pos + 2]
            fs = _fields(cfg)
            need = 2 + sum(_SIZE[d] for _, d in fs)
            if len(buf) - pos < need:
                if eof:
                    break
                buf = buf[pos:]; base += pos; pos = 0
                more = f.read(CH)
                if not more:
                    eof = True
                buf += more
                if len(buf) < need:
                    break
                continue
            q = pos + 2
            vals = []
            for logic, dtp in fs:
                sz = _SIZE[dtp]
                raw = int.from_bytes(buf[q:q + sz], 'big', signed=(logic == ASIS))
                q += sz
                vals.append(-raw if logic == MINUS else raw)
            pos = q
            t += vals[0]; a += vals[1]; bd += vals[2]; v += vals[3]
        ts.append(t); asks.append(a); bids.append(bd); vols.append(v)
        n += 1
        if progress and n % progress == 0:
            print(f"  {n:,} ticks  {dt.datetime.fromtimestamp(t/1000, dt.UTC):%Y-%m-%d}", flush=True)
        if len(ts) >= batch:
            yield (np.array(ts, dtype=np.int64), np.array(asks, dtype=np.int64),
                   np.array(bids, dtype=np.int64), np.array(vols, dtype=np.int64))
            ts, asks, bids, vols = [], [], [], []
        if max_records and n >= max_records:
            break
    f.close()
    if ts:
        yield (np.array(ts, dtype=np.int64), np.array(asks, dtype=np.int64),
               np.array(bids, dtype=np.int64), np.array(vols, dtype=np.int64))


def calibrate_decimals(ts, ask_i, ref_time_s, ref_close, verbose=True):
    """Determine the price scale by matching against a KNOWN reference series.

    SQX keeps `decimals` in metadata outside the .dat, and it cannot be inferred from the
    file alone: 1216010000 is a plausible price at 10^3 (1216010), 10^5 (12160.1) or 10^6
    (1216.01), and nothing in the bytes distinguishes them. Guessing here would be exactly
    the kind of silent, plausible-looking error this project keeps getting burned by - a
    100x price scale error would not crash anything, it would just quietly rescale every
    ATR-normalised feature downstream.

    So instead: take a reference series for the same instrument (the MT5 rate exports under
    Common\\Files\\Warrior_EA\\Research), find the tick nearest each reference bar time, and
    pick the power of ten minimising median relative error. A correct scale lands at ~0;
    every wrong one is off by a factor of ten or more, so the decision is never marginal.
    """
    ref_ms = (ref_time_s * 1000).astype(np.int64)
    lo, hi = ts[0], ts[-1]
    m = (ref_ms >= lo) & (ref_ms <= hi)
    if m.sum() < 50:
        raise ValueError(f"only {int(m.sum())} overlapping reference bars - cannot calibrate")
    rt, rc = ref_ms[m], ref_close[m]
    idx = np.searchsorted(ts, rt).clip(1, len(ts) - 1)
    got = ask_i[idx].astype(float)
    best, best_err = None, np.inf
    for d in range(0, 11):
        err = float(np.median(np.abs(got / (10.0 ** d) - rc) / np.maximum(np.abs(rc), 1e-9)))
        if verbose:
            print(f"    decimals={d:<3} median rel.err {err:12.6f}")
        if err < best_err:
            best, best_err = d, err
    if best_err > 0.05:
        raise ValueError(f"best scale (decimals={best}) still off by {best_err:.1%} - "
                         "reference series may be a different instrument")
    return best, best_err


def to_frame(path, max_records=None, decimals=None, volume_constant=VOLUME_CONSTANT,
             progress=None, ref=None):
    """decimals: pass explicitly, or pass ref=(time_s, close) to calibrate against a known
    series. There is no safe default - see calibrate_decimals()."""
    ts, ai, bi, vi = decode(path, max_records, progress)
    if decimals is None:
        if ref is None:
            raise ValueError("pass decimals=... or ref=(time_s, close) - see calibrate_decimals()")
        decimals, _ = calibrate_decimals(ts, ai, ref[0], ref[1])
    scale = 10.0 ** decimals
    ask = ai / scale
    bid = bi / scale
    vol = np.where(vi == NO_VOLUME, np.nan, vi / volume_constant)
    return ts, bid, ask, vol, decimals


if __name__ == '__main__':
    # usage: python sqx.py <file.dat> [max_records] [--ref SYMBOL | --decimals N]
    sys.stdout.reconfigure(encoding='utf-8', errors='replace')
    path = sys.argv[1]
    lim = int(sys.argv[2]) if len(sys.argv) > 2 and not sys.argv[2].startswith('-') else None
    dec_arg, ref_sym = None, None
    if '--decimals' in sys.argv:
        dec_arg = int(sys.argv[sys.argv.index('--decimals') + 1])
    if '--ref' in sys.argv:
        ref_sym = sys.argv[sys.argv.index('--ref') + 1]
    ref = None
    if ref_sym:
        from kit import load_rates
        rt, ro, rh, rl, rc, rv, rs = load_rates(ref_sym, 16385)
        ref = (rt, rc)
        print(f"calibrating scale against {ref_sym} H1 closes:")
    ts, bid, ask, vol, dec = to_frame(path, max_records=lim, decimals=dec_arg,
                                      progress=5000000, ref=ref)
    print(f"\ndecoded {len(ts):,} ticks   decimals={dec}")
    print(f"span  {dt.datetime.fromtimestamp(ts[0]/1000, dt.UTC)}"
          f"  ->  {dt.datetime.fromtimestamp(ts[-1]/1000, dt.UTC)}")
    print(f"bid   min {np.nanmin(bid):.5f}  max {np.nanmax(bid):.5f}")
    spread = ask - bid
    print(f"spread median {np.nanmedian(spread):.5f}   negative: {(spread<0).sum()}")
    print(f"time monotonic: {bool((np.diff(ts)>=0).all())}   "
          f"max gap {np.max(np.diff(ts))/1000/3600:.1f} h")
    print(f"volume: median {np.nanmedian(vol)}  nan {int(np.isnan(vol).sum()):,}")
    print("\nfirst 5:")
    for i in range(5):
        print(f"  {dt.datetime.fromtimestamp(ts[i]/1000, dt.UTC)}  "
              f"bid {bid[i]:.5f}  ask {ask[i]:.5f}  vol {vol[i]}")