본문으로 건너뛰기
라이브러리 문서 전체

경제적 가설과 데이터 시점에 맞춘 금융 특성 설계

코드 Machine Learning for Trading

요약

이 장은 트레이딩 아이디어를 문서화된 특성 명세로 바꾸는 프레임워크를 제시합니다. 연구자는 특성의 시간 범위를 의도한 의사결정에 맞추고, 구동 요인 가설을 밝히며, 예측 신호와 맥락을 설명하는 상태 변수를 구분해야 합니다. 기준 관점, 표현 방식, 집계 방법은 경제적 주장에 따라 정해야 합니다. 가설을 바꾸는 변환과 주로 노이즈를 조절하는 변환도 구별해야 합니다. 이 장은 가격·거래량 특성, 시장 미시구조, 상품 간 관계, 파생상품 내재 지표, 펀더멘털, 거시경제 데이터, 달력 변수를 살펴봅니다.

더 빠른 신호와 느린 상태 변수를 게이팅, 스케일링, 조건부 변형으로 결합하면서 상호작용이 정보를 더하는지 확인하는 방법을 강조합니다. 보고 지연과 수정이 잘못된 증거를 만들 수 있으므로 펀더멘털과 거시 데이터는 시점별 이용 가능성을 확인해야 합니다. 통제된 변형, 같은 계열 내 중복 제거, 견고성 검사로 특성 탐색도 체계적으로 진행해야 합니다. 이 장은 하나의 전략이나 단일 성과 결과를 제시하는 대신 방법과 실패 요인을 설명합니다. 각 특성군은 적절한 데이터, 시점 가정, 검증, 비용을 고려한 평가가 여전히 필요합니다.

핵심 아이디어

  • 특성을 설계할 때는 시간 범위의 정합성, 구동 요인 가설, 특성의 역할을 먼저 명확히 합니다.
  • 기준 관점, 표현 방식, 집계 방법을 어떻게 선택하는지에 따라 특성의 경제적 의미가 달라질 수 있습니다.
  • 느리게 변하는 펀더멘털과 거시 변수는 발표 시차와 수정 내역을 반영해 시점에 맞춰야 합니다.
  • 신호와 상태 변수의 상호작용은 조건부 정보를 더할 수 있지만, 검토할 선택지도 늘어납니다.
  • 중복 제거, 통제된 변형, 이벤트 연구, 견고성 검사는 신뢰할 수 있는 특성 선택에 도움이 됩니다.

태그

전문
# artifact_audit.py


```py
#!/usr/bin/env python3
"""Report registered fits whose named feature artifact is not on disk.

A training identity carries ``computation.feature_artifacts.<role>.sha256``, the hash of
the artifact the fit read. Nothing else checks the file is still there, so a fit whose
input is gone resolves, reads as healthy, and is a record of a result rather than a
reproducible one.

**The scope this walks is printed, and that is not decoration.** An earlier run of an
uncommitted version of this audit reported ``crypto_perps_funding/financial`` as missing
for 144 fits while the file sat at the canonical path under its registered hash, untouched
for two months (ml4t/agent-workspace#1176). A case study's ``features/`` is a symlink into
the artifacts root in every checkout that has one, so an audit run from a throwaway
worktree - or one that does not resolve the link - hashes a different tree than the fits
read and reports absence with full confidence. Every path below is resolved and printed,
and a resolved path outside the artifacts root is a hard error rather than a quiet
mis-scope.
"""

from __future__ import annotations

import argparse
import hashlib
import json
import sqlite3
import sys
from collections import Counter
from pathlib import Path

REPO_ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(REPO_ROOT))

from utils.paths import get_case_study_dir  # noqa: E402

CASE_STUDIES = (
    "cme_futures",
    "crypto_perps_funding",
    "etfs",
    "fx_pairs",
    "nasdaq100_microstructure",
    "sp500_equity_option_analytics",
    "sp500_options",
    "us_equities_panel",
    "us_firm_characteristics",
)
ARTIFACT_DIRS = ("features", "labels")


def sha256_of(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for block in iter(lambda: handle.read(1 << 20), b""):
            digest.update(block)
    return digest.hexdigest()


def registered_artifacts(spec_json: str) -> list[tuple[str, str, int | None]]:
    """``(role, sha256, size)`` for one spec, across both shapes the registry holds.

    A dict keyed by role, and a list of ``{role, sha256}`` whose hashes may carry a
    ``sha256:`` prefix. Matching on the hash alone reaches both, but the role and size are
    what make a report readable, so both shapes are parsed rather than regex-scraped.
    """
    artifacts = json.loads(spec_json).get("computation", {}).get("feature_artifacts")
    out: list[tuple[str, str, int | None]] = []
    if isinstance(artifacts, dict):
        for role, entry in artifacts.items():
            if isinstance(entry, dict) and entry.get("sha256"):
                out.append((role, str(entry["sha256"]).removeprefix("sha256:"), entry.get("size")))
    elif isinstance(artifacts, list):
        for entry in artifacts:
            if isinstance(entry, dict) and entry.get("sha256"):
                out.append(
                    (
                        str(entry.get("role", "?")),
                        str(entry["sha256"]).removeprefix("sha256:"),
                        entry.get("size"),
                    )
                )
    return out


def walk(
    case_study: str, artifacts_root: Path | None
) -> tuple[dict[str, Path], list[str], list[str]]:
    """Hash every parquet the case study's artifact directories hold, after resolving them.

    The third return is the directories that were not there. An absent one is not a small
    gap in an otherwise good answer: every hash it would have contributed is absent from
    ``on_disk``, so every fit that read it is counted as naming a missing artifact. The
    caller has to know, because the two readings are opposite and the output is identical.
    """
    case_dir = get_case_study_dir(case_study)
    on_disk: dict[str, Path] = {}
    scope: list[str] = []
    absent: list[str] = []
    for name in ARTIFACT_DIRS:
        declared = case_dir / name
        if not declared.exists():
            scope.append(f"    {name}/  ABSENT at {declared}")
            absent.append(f"{case_study}/{name} at {declared}")
            continue
        resolved = declared.resolve()
        scope.append(f"    {name}/  -> {resolved}")
        if artifacts_root is not None and artifacts_root not in resolved.parents:
            raise SystemExit(
                f"{case_study}/{name} resolves to {resolved}, which is not under "
                f"{artifacts_root}. Hashing it would compare the registry's fits against a "
                f"tree they never read. Run from a checkout whose artifact directories link "
                f"into the artifacts root, or pass --artifacts-root to name the tree you mean."
            )
        for parquet in sorted(resolved.rglob("*.parquet")):
            on_disk.setdefault(sha256_of(parquet), parquet)
    return on_disk, scope, absent


def default_artifacts_root(case_studies: tuple[str, ...]) -> Path | None:
    """The tree the artifact directories actually live in, taken from the first that exists.

    Derived from a RESOLVED artifact directory, never from the case-study directory: in a
    checkout those are two different trees, because ``case_studies/<cs>/features`` is a
    symlink into the artifacts root and its unresolved parent is the repository.
    """
    for case_study in case_studies:
        case_dir = get_case_study_dir(case_study)
        for name in ARTIFACT_DIRS:
            declared = case_dir / name
            if declared.exists():
                # <root>/<case study>/<features|labels>
                return declared.resolve().parents[1]
    return None


def audit(
    case_studies: tuple[str, ...], artifacts_root: Path | None
) -> tuple[int, list[str], list[str]]:
    """Returns the missing-fit count, the case studies skipped, and the directories unseen.

    The second and third are not details. A partial audit that prints "0 fits name an
    artifact that is not on disk" is indistinguishable from a clean one, and a checkout is
    missing a registry whenever its gitignored ``run_log`` symlink was never created -
    which is the normal state of most worktrees, not an exception.

    The third exists because the second was not enough. An absent *artifact directory*
    inside an audited case study used to print one ABSENT line in the middle of the scope
    block and then contribute every one of that case study's fits to a confident total.
    Measured 2026-09-14 on a checkout whose ``us_equities_panel`` had ``run_log`` but no
    ``features/``: the audit reported 393 fits naming missing artifacts and exited 1, with
    no PARTIAL, while ``financial.parquet`` sat on disk at exactly the 4,478,156,899 bytes
    the MISSING line quoted. A case study that could not be fully seen is now not counted
    at all, because there is no way to tell its real findings from its blind ones.
    """
    total_missing = 0
    unaudited: list[str] = []
    unseen: list[str] = []
    for case_study in case_studies:
        db = get_case_study_dir(case_study) / "run_log" / "registry.db"
        print(f"\n{case_study}")
        if not db.exists():
            print(f"    NOT AUDITED: no registry at {db}")
            unaudited.append(case_study)
            continue
        on_disk, scope, absent = walk(case_study, artifacts_root)
        for line in scope:
            print(line)
        if absent:
            print(
                f"    NOT AUDITED: {len(absent)} artifact directory(ies) absent, so every "
                f"fit here would read as missing. Nothing from this case study is counted."
            )
            unseen.extend(absent)
            continue
        print(f"    {len(on_disk)} distinct parquet hashed")

        connection = sqlite3.connect(f"file:{db}?mode=ro", uri=True)
        try:
            rows = connection.execute("SELECT training_hash, spec_json FROM training_runs")
            missing: Counter[tuple[str, str, int | None]] = Counter()
            fits = 0
            for _, spec_json in rows:
                fits += 1
                for role, sha, size in registered_artifacts(spec_json):
                    if sha not in on_disk:
                        missing[(role, sha, size)] += 1
        finally:
            connection.close()

        if not missing:
            print(f"    {fits} fits, every named artifact present")
            continue
        for (role, sha, size), count in sorted(missing.items(), key=lambda kv: -kv[1]):
            size_text = f"{size:,} bytes" if size else "size not recorded"
            print(f"    MISSING  {role:<14} {sha[:16]}…  {size_text}  named by {count} fits")
            total_missing += count
    return total_missing, unaudited, unseen


def main() -> int:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--case-study", action="append", choices=CASE_STUDIES)
    parser.add_argument(
        "--artifacts-root",
        type=Path,
        help=(
            "Tree every artifact directory must resolve under. Defaults to the parent of the "
            "first case study's resolved directory. Pass 'none' to disable the check, which "
            "is how you audit a deliberately isolated tree."
        ),
    )
    args = parser.parse_args()
    selected = tuple(args.case_study) if args.case_study else CASE_STUDIES

    if args.artifacts_root is not None and str(args.artifacts_root) == "none":
        artifacts_root = None
    elif args.artifacts_root is not None:
        artifacts_root = args.artifacts_root.resolve()
    else:
        artifacts_root = default_artifacts_root(selected)
        if artifacts_root is None:
            raise SystemExit(
                "no case study has a features/ or labels/ directory, so there is no tree to "
                "audit. Pass --artifacts-root to name one explicitly."
            )

    print(f"artifacts root: {artifacts_root if artifacts_root else 'UNCHECKED'}")
    total, unaudited, unseen = audit(selected, artifacts_root)
    if unaudited or unseen:
        # Deliberately no total. A number printed here is read as the answer however it is
        # qualified, and over an incomplete scope it is not one.
        print("\nPARTIAL: this audit did not see the whole tree, so it reports no count.")
        if unaudited:
            print(
                f"  {len(unaudited)} of {len(selected)} case studies had no registry: "
                f"{', '.join(unaudited)}"
            )
        if unseen:
            print(f"  {len(unseen)} artifact directory(ies) absent:")
            for line in unseen:
                print(f"    {line}")
        print("  Run from a checkout where every case study links into the artifacts root.")
        return 2
    print(f"\n{total} fits name an artifact that is not on disk")
    return 1 if total else 0


if __name__ == "__main__":
    raise SystemExit(main())

```

출처의 라이선스에 따라 출처를 표시하고 전문을 공개합니다. 라이선스: MIT

이 요약은 원문을 바탕으로 Stratmill의 리서치 에이전트가 작성했으며, 원문을 복사한 것이 아닙니다.