MacroLens / code /validate_all.py
itouchz's picture
Add files using upload-large-folder tool
5995ef5 verified
"""Expanded validation suite for MacroLens benchmark outputs.
Runs >=130 atomic checks across 20 sections. Each check is either PASS, WARN,
or FAIL. Exits non-zero if any FAIL. Designed to be run after the pipeline
completes a clean rerun — catches contamination, regressions, and data
quality issues that the 62-check notebook misses.
Usage:
uv run --env-file .env python -m projects.agent_builder.scripts.whatif_bench.validate_all
"""
from __future__ import annotations
import json
import sys
from dataclasses import dataclass, field
from pathlib import Path
import numpy as np
import pandas as pd
from projects.agent_builder.scripts.whatif_bench import config
ROOT = Path(config.DATA_DIR) if hasattr(config, "DATA_DIR") else Path(__file__).parent / "data_small_caps"
from functools import lru_cache
@lru_cache(maxsize=1)
def _xbrl_required_map() -> dict[str, bool]:
"""Per-ticker XBRL applicability from SEC EDGAR's companyfacts cache.
Empty SEC responses (no XBRL) are exactly 52 bytes; real ones are 16KB+.
Use file size as the discriminator — no JSON parsing, ~50ms for 4400 files.
Cached at module level so the validator's 3 granularity checks share the
same dict instead of re-reading 600 MB of JSON three times.
"""
raw_dir = ROOT / "xbrl" / "raw"
if not raw_dir.exists():
return {}
out: dict[str, bool] = {}
for f in raw_dir.glob("*.json"):
# SEC's empty/missing companyfacts response is a tiny stub.
out[f.stem] = f.stat().st_size > 100
return out
@dataclass
class Report:
pass_count: int = 0
warn_count: int = 0
fail_count: int = 0
lines: list[str] = field(default_factory=list)
def add(self, name: str, ok: bool, detail: str = "", warn: bool = False) -> None:
if ok:
self.pass_count += 1
self.lines.append(f" PASS {name}{(' — ' + detail) if detail else ''}")
elif warn:
self.warn_count += 1
self.lines.append(f" WARN {name}{detail}")
else:
self.fail_count += 1
self.lines.append(f" FAIL {name}{detail}")
@property
def total(self) -> int:
return self.pass_count + self.warn_count + self.fail_count
def section(title: str, r: Report) -> None:
r.lines.append("")
r.lines.append(f"=== {title} ===")
def check_universe(r: Report) -> None:
section("1. Universe Integrity (8 checks)", r)
p = ROOT / "universe" / "benchmark_universe.csv"
r.add("1.1 universe file exists", p.exists(), str(p))
if not p.exists():
return
u = pd.read_csv(p)
r.add("1.2 ticker count == 4416", len(u) == 4416, f"got {len(u)}")
r.add("1.3 no duplicate tickers", u["ticker"].is_unique, f"{u['ticker'].duplicated().sum()} dupes")
r.add("1.4 no null tickers", u["ticker"].notna().all())
has_mv = "market_value" in u.columns
if has_mv:
mega = u[(u["market_value"] > 7.4e9) & ~u.get("in_russell_2000", False)]
r.add("1.5 no mega-caps outside R2K (> $7.4B)", mega.empty, f"{len(mega)} found")
in_r2k = u.get("in_russell_2000", pd.Series(dtype=bool)).sum()
r.add("1.6 R2K membership populated", in_r2k > 1500, f"{in_r2k} in R2K")
# Real S&P 600 membership requires the in_sp_smallcap_600 flag to be
# set by the collector. Counting source=='IJR' undercounts because
# tickers present in both R2K and SP600 are tagged source='IWM'. Fail
# here until the collector populates a proper membership flag.
in_sp600 = int(u.get("in_sp_smallcap_600", pd.Series(dtype=bool)).sum()) \
if "in_sp_smallcap_600" in u.columns else 0
r.add("1.7 S&P 600 membership populated", in_sp600 > 400,
f"{in_sp600} in SP600 (need collector to populate `in_sp_smallcap_600` flag from full IJR holdings)")
r.add("1.8 at least one source flag per ticker", True, "skipped (schema-dependent)", warn=True)
def check_panel_schema(r: Report, gran: str) -> pd.DataFrame | None:
section(f"2.{gran} Panel Schema ({gran}, 10 checks)", r)
p = ROOT / "processed" / gran / "panel.parquet"
r.add(f"2.{gran}.1 panel file exists", p.exists(), str(p))
if not p.exists():
return None
df = pd.read_parquet(p)
expected_rows = {"daily": 4_841_094, "weekly": 1_009_314, "monthly": 232_483}[gran]
r.add(f"2.{gran}.2 row count == {expected_rows}", len(df) == expected_rows, f"got {len(df)}")
r.add(f"2.{gran}.3 column count == 137", len(df.columns) == 137, f"got {len(df.columns)}")
for col in ("ticker", "date", "close", "adj_close", "sector", "industry", "exchange"):
r.add(f"2.{gran}.4 col {col} present", col in df.columns)
r.add(f"2.{gran}.5 primary key (ticker,date) unique", not df.duplicated(["ticker", "date"]).any(),
f"{df.duplicated(['ticker','date']).sum()} dupes")
r.add(f"2.{gran}.6 date is datetime dtype", pd.api.types.is_datetime64_any_dtype(df["date"]), str(df["date"].dtype))
r.add(f"2.{gran}.7 ticker is string", df["ticker"].dtype == object or str(df["ticker"].dtype).startswith("string"))
r.add(f"2.{gran}.8 date range starts 2021-01-01", df["date"].min() >= pd.Timestamp("2021-01-01"),
f"min={df['date'].min()}")
# Weekly resample rounds to Friday; allow up to 2026-04-10 buffer
max_allowed = pd.Timestamp("2026-04-10") if gran == "weekly" else pd.Timestamp("2026-04-01")
r.add(f"2.{gran}.9 date range ends <= {max_allowed.date()}", df["date"].max() <= max_allowed,
f"max={df['date'].max()}")
r.add(f"2.{gran}.10 unique ticker count == 4416", df["ticker"].nunique() == 4416, f"got {df['ticker'].nunique()}")
return df
def check_prices(r: Report, df: pd.DataFrame, gran: str) -> None:
section(f"3.{gran} Price Invariants ({gran}, 8 checks)", r)
neg_adj = (df["adj_close"] < 0).sum()
r.add(f"3.{gran}.1 no negative adj_close", neg_adj == 0, f"{neg_adj} rows")
hilo = (df["high"] < df["low"]).sum()
r.add(f"3.{gran}.2 high >= low", hilo == 0, f"{hilo} rows")
ho = (df["high"] < df["open"]).sum()
r.add(f"3.{gran}.3 high >= open", ho == 0, f"{ho} rows")
hc = (df["high"] < df["close"]).sum()
r.add(f"3.{gran}.4 high >= close", hc == 0, f"{hc} rows")
lo = (df["low"] > df["open"]).sum()
r.add(f"3.{gran}.5 low <= open", lo == 0, f"{lo} rows")
lc = (df["low"] > df["close"]).sum()
r.add(f"3.{gran}.6 low <= close", lc == 0, f"{lc} rows")
nvol = (df["volume"] < 0).sum()
r.add(f"3.{gran}.7 no negative volume", nvol == 0, f"{nvol} rows")
nclose = df["close"].isna().sum()
r.add(f"3.{gran}.8 close coverage > 99%", nclose / len(df) < 0.01, f"{nclose} NaN ({100*nclose/len(df):.2f}%)")
def check_xbrl_balance(r: Report, df: pd.DataFrame, gran: str) -> None:
section(f"4.{gran} XBRL / Balance Equation ({gran}, 10 checks)", r)
for col in ("stmt_total_assets", "stmt_total_liabilities", "stmt_total_equity", "stmt_revenue"):
r.add(f"4.{gran}.col {col} present", col in df.columns)
if "stmt_total_assets" not in df.columns:
return
sub = df.dropna(subset=["stmt_total_assets", "stmt_total_liabilities", "stmt_total_equity"])
if len(sub) > 0:
diff = (sub["stmt_total_assets"] - (sub["stmt_total_liabilities"] + sub["stmt_total_equity"])).abs()
rel = diff / sub["stmt_total_assets"].abs().clip(lower=1)
bad = (rel > 0.01).sum()
r.add(f"4.{gran}.1 balance equation A=L+E within 1% (all rows)", bad == 0,
f"{bad}/{len(sub)} rows fail ({100*bad/len(sub):.2f}%)")
neg_rev = (df["stmt_revenue"] < 0).sum()
r.add(f"4.{gran}.2 no negative revenue (after recovery)", neg_rev == 0, f"{neg_rev} rows")
neg_assets = (df["stmt_total_assets"] <= 0).sum()
r.add(f"4.{gran}.3 assets > 0 (after recovery)", neg_assets == 0, f"{neg_assets} rows")
neg_liab = (df["stmt_total_liabilities"] < 0).sum()
r.add(f"4.{gran}.4 liabilities >= 0 (after recovery)", neg_liab == 0, f"{neg_liab} rows")
rev_cov = df["stmt_revenue"].notna().sum() / len(df)
r.add(f"4.{gran}.5 revenue coverage > 80%", rev_cov > 0.80, f"{100*rev_cov:.1f}%")
# Applicability-aware coverage, grounded in SEC EDGAR ground truth.
# A ticker is "XBRL-required" iff its companyfacts JSON cached at
# data_small_caps/xbrl/raw/{TICKER}.json contains a non-empty
# `facts.us-gaap` dict. Closed-end funds, foreign 6-K filers, royalty
# trusts, FDIC-only banks return empty/missing us-gaap → exempt.
# Per-row applicability = (ticker XBRL-required) AND (date >= ticker's
# first SEC filing date). Threshold: effective coverage > 99%.
xbrl_required = _xbrl_required_map()
# Tickers NOT in raw dir are unknown — treat as not-required (conservative)
applicable_ticker_mask = df["ticker"].map(xbrl_required).fillna(False).astype(bool)
# First SEC filing date per ticker — vectorized via map (NOT df.apply)
if "nearest_filing_date" in df.columns:
first_filing_date = (
df[df["nearest_filing_date"].notna()]
.groupby("ticker")["nearest_filing_date"].min()
)
first_per_row = df["ticker"].map(first_filing_date)
date_ge_first = (df["date"] >= first_per_row).fillna(False)
else:
date_ge_first = pd.Series(False, index=df.index)
applicable = applicable_ticker_mask & date_ge_first
col = df["stmt_total_assets"]
n_applicable = int(applicable.sum())
n_filled = int((col.notna() & applicable).sum())
n_required_tickers = sum(1 for v in xbrl_required.values() if v)
n_exempt_tickers = sum(1 for v in xbrl_required.values() if not v)
eff = n_filled / max(1, n_applicable)
naive = col.notna().sum() / len(df)
r.add(
f"4.{gran}.6 total_assets effective coverage > 99% (SEC-EDGAR-grounded)",
eff > 0.99,
f"effective={100*eff:.2f}% (filled {n_filled:,}/applicable {n_applicable:,}); "
f"naive={100*naive:.1f}%; "
f"{n_required_tickers} tickers XBRL-required, {n_exempt_tickers} exempt per SEC EDGAR",
)
def check_derived(r: Report, df: pd.DataFrame, gran: str) -> None:
section(f"5.{gran} Derived Metrics ({gran}, 8 checks)", r)
if "derived_market_cap" in df.columns:
mc = df["derived_market_cap"].dropna()
r.add(f"5.{gran}.1 market_cap <= $100B ceiling", (mc <= 100e9).all(), f"max={mc.max():.2e}")
r.add(f"5.{gran}.2 market_cap > 0", (mc > 0).all(), f"min={mc.min():.2e}")
if "derived_pe" in df.columns:
pe = df["derived_pe"].dropna()
r.add(f"5.{gran}.3 PE values finite", np.isfinite(pe).all())
r.add(f"5.{gran}.4 PE > 0 (loss-making set to NaN)", (pe > 0).all(), f"min={pe.min()}", warn=True)
if "derived_pb" in df.columns:
pb = df["derived_pb"].dropna()
r.add(f"5.{gran}.5 PB finite", np.isfinite(pb).all())
if "derived_gross_margin" in df.columns:
gm = df["derived_gross_margin"].dropna()
r.add(f"5.{gran}.6 gross_margin in [-5, 5]", ((gm >= -5) & (gm <= 5)).all(),
f"range=[{gm.min():.2f},{gm.max():.2f}]", warn=not ((gm >= -5) & (gm <= 5)).all())
if "derived_wacc" in df.columns:
wacc = df["derived_wacc"].dropna()
r.add(f"5.{gran}.7 wacc in [0, 1]", ((wacc >= 0) & (wacc <= 1)).all(),
f"range=[{wacc.min():.3f},{wacc.max():.3f}]", warn=True)
r.add(f"5.{gran}.8 derived_* col count >= 15", sum(1 for c in df.columns if c.startswith("derived_")) >= 15)
def check_macro(r: Report, df: pd.DataFrame, gran: str) -> None:
section(f"6.{gran} Macro Data ({gran}, 6 checks)", r)
fred_cols = [c for c in df.columns if c.startswith("fred_")]
eia_cols = [c for c in df.columns if c.startswith("eia_")]
r.add(f"6.{gran}.1 fred_* col count >= 40", len(fred_cols) >= 40, f"got {len(fred_cols)}")
r.add(f"6.{gran}.2 eia_* col count >= 5", len(eia_cols) >= 5, f"got {len(eia_cols)}")
if "fred_vix" in df.columns:
vix = df["fred_vix"].dropna()
r.add(f"6.{gran}.3 VIX in [5, 100]", ((vix >= 5) & (vix <= 100)).all(),
f"range=[{vix.min():.1f},{vix.max():.1f}]", warn=True)
if fred_cols:
nan_rate = df[fred_cols].isna().mean().mean()
r.add(f"6.{gran}.4 fred NaN rate < 20%", nan_rate < 0.20, f"{100*nan_rate:.1f}%")
if "fred_dgs10" in df.columns:
yld = df["fred_dgs10"].dropna()
r.add(f"6.{gran}.5 10Y yield in [-1, 10]", ((yld >= -1) & (yld <= 10)).all(),
f"range=[{yld.min():.2f},{yld.max():.2f}]", warn=True)
oil_cols = [c for c in df.columns if ("crude" in c.lower() and "spot" in c.lower())
or c == "fred_DCOILWTICO"]
r.add(f"6.{gran}.6 oil price col present", len(oil_cols) > 0, f"found: {oil_cols[:2]}")
def check_filing_context(r: Report, df: pd.DataFrame, gran: str) -> None:
section(f"7.{gran} Filing Context ({gran}, 5 checks)", r)
for col in ("nearest_filing_type", "nearest_filing_date", "days_since_filing"):
r.add(f"7.{gran}.col {col} present", col in df.columns)
if "days_since_filing" in df.columns:
d = df["days_since_filing"].dropna()
r.add(f"7.{gran}.1 days_since_filing >= 0", (d >= 0).all(), f"min={d.min()}")
if "nearest_filing_date" in df.columns and "date" in df.columns:
sub = df.dropna(subset=["nearest_filing_date"])
r.add(f"7.{gran}.2 nearest_filing_date <= date", (sub["nearest_filing_date"] <= sub["date"]).all())
if "nearest_filing_type" in df.columns:
types = df["nearest_filing_type"].dropna().unique()
# Every explicitly collected form type. 'other' is NOT accepted —
# if any filings land in 'other', the classifier needs to be
# extended to handle them properly.
expected = {
"10-K", "10-K/A", "10-Q", "10-Q/A", "8-K", "20-F", "40-F",
"N-CSR", "N-CSRS", "6-K", "DEF 14A", "S-1", "11-K",
}
unexpected = set(types) - expected
r.add(f"7.{gran}.3 filing types in expected set", len(unexpected) == 0,
f"unexpected: {unexpected}")
def check_split_integrity(r: Report, gran: str) -> None:
section(f"8.{gran} Train/Test Split ({gran}, 6 checks)", r)
bp = ROOT / "benchmark" / gran
train_p = bp / "panel_train.parquet"
test_p = bp / "panel_test.parquet"
r.add(f"8.{gran}.1 train parquet exists", train_p.exists())
r.add(f"8.{gran}.2 test parquet exists", test_p.exists())
if not (train_p.exists() and test_p.exists()):
return
tr = pd.read_parquet(train_p, columns=["ticker", "date"])
te = pd.read_parquet(test_p, columns=["ticker", "date"])
tr_keys = set(zip(tr["ticker"], tr["date"]))
te_keys = set(zip(te["ticker"], te["date"]))
overlap = tr_keys & te_keys
r.add(f"8.{gran}.3 no (ticker,date) overlap train/test", len(overlap) == 0, f"{len(overlap)} keys overlap")
r.add(f"8.{gran}.4 train.date.max() < test.date.min() (temporal)",
tr["date"].max() < te["date"].min(), f"tr_max={tr['date'].max()} te_min={te['date'].min()}")
ratio = len(tr) / (len(tr) + len(te))
r.add(f"8.{gran}.5 train/(train+test) ratio in [0.65, 0.75]", 0.65 <= ratio <= 0.75, f"ratio={ratio:.3f}")
r.add(f"8.{gran}.6 both splits non-empty", len(tr) > 0 and len(te) > 0)
def check_scenarios(r: Report, gran: str) -> None:
section(f"9.{gran} Scenarios ({gran}, 7 checks)", r)
p = ROOT / "benchmark" / gran / "scenarios.parquet"
r.add(f"9.{gran}.1 scenarios file exists", p.exists())
if not p.exists():
return
s = pd.read_parquet(p)
expected = {"daily": 1130, "weekly": None, "monthly": None}
if expected[gran] is not None:
r.add(f"9.{gran}.2 scenario count == {expected[gran]}", len(s) == expected[gran], f"got {len(s)}")
r.add(f"9.{gran}.3 scenarios non-empty", len(s) > 0)
r.add(f"9.{gran}.4 event_date col present", "event_date" in s.columns)
r.add(f"9.{gran}.5 event_type col present", "event_type" in s.columns)
if "event_type" in s.columns:
n_types = s["event_type"].nunique()
r.add(f"9.{gran}.6 event_type variety >= 20", n_types >= 20, f"got {n_types}")
if "event_date" in s.columns:
ed = pd.to_datetime(s["event_date"])
r.add(f"9.{gran}.7 events within panel window [2021,2026]",
(ed >= pd.Timestamp("2021-01-01")).all() and (ed <= pd.Timestamp("2026-04-01")).all())
def check_benchmark_files(r: Report, gran: str) -> None:
section(f"10.{gran} Benchmark Task Files ({gran}, 15 checks)", r)
bp = ROOT / "benchmark" / gran
task_files = [
"panel_train.parquet", "panel_test.parquet", "task_definition.json", "metadata.json",
"filing_corpus.parquet", "scenarios.parquet",
"valuation_inputs.parquet", "valuation_ground_truth.parquet",
"generation_inputs.parquet", "generation_ground_truth.parquet",
"scenario_forecast_ground_truth.parquet",
"private_valuation_inputs.parquet", "private_valuation_ground_truth.parquet",
"generator_eval_inputs.parquet", "generator_eval_ground_truth.parquet",
# Task F (RE-Val) artifacts -- written by build_valuation_tasks.py
# _build_task_f. Previously missing here, so a half-built T7 would
# pass validation silently.
"re_train_properties.parquet", "re_eval_inputs.parquet",
"re_eval_ground_truth.parquet",
]
for i, f in enumerate(task_files, 1):
r.add(f"10.{gran}.{i} {f}", (bp / f).exists(), str(bp / f))
def check_valuation_holdout(r: Report, gran: str) -> None:
section(f"11.{gran} Valuation Holdout Integrity ({gran}, 4 checks)", r)
bp = ROOT / "benchmark" / gran
vi = bp / "valuation_inputs.parquet"
vg = bp / "valuation_ground_truth.parquet"
if not (vi.exists() and vg.exists()):
r.add(f"11.{gran}.1 valuation files present", False, "missing inputs or ground truth")
return
v_in = pd.read_parquet(vi)
v_gt = pd.read_parquet(vg)
r.add(f"11.{gran}.1 valuation non-empty", len(v_in) > 0 and len(v_gt) > 0)
if "ticker" in v_in.columns and "ticker" in v_gt.columns:
r.add(f"11.{gran}.2 inputs tickers == ground_truth tickers",
set(v_in["ticker"]) == set(v_gt["ticker"]),
f"{len(set(v_in['ticker']) ^ set(v_gt['ticker']))} diff")
r.add(f"11.{gran}.3 ground_truth has target col", any(c for c in v_gt.columns if "market_cap" in c.lower() or "target" in c.lower()),
f"cols={list(v_gt.columns)[:5]}", warn=True)
if "ticker" in v_gt.columns:
n = v_gt["ticker"].nunique()
r.add(f"11.{gran}.4 holdout tickers ~ 30% of universe", 1000 < n < 1700, f"{n} holdout tickers")
def check_cross_granularity(r: Report) -> None:
section("12. Cross-Granularity Consistency (4 checks)", r)
panels = {}
for gran in ("daily", "weekly", "monthly"):
p = ROOT / "processed" / gran / "panel.parquet"
if p.exists():
panels[gran] = pd.read_parquet(p, columns=["ticker", "date"])
if len(panels) < 2:
r.add("12.1 all granularities present", False, f"got {list(panels)}")
return
r.add("12.1 all 3 granularities present", len(panels) == 3)
tickers = [set(p["ticker"]) for p in panels.values()]
r.add("12.2 same ticker set across granularities", all(t == tickers[0] for t in tickers))
if "daily" in panels and "weekly" in panels:
r.add("12.3 weekly rows < daily rows", len(panels["weekly"]) < len(panels["daily"]))
if "weekly" in panels and "monthly" in panels:
r.add("12.4 monthly rows < weekly rows", len(panels["monthly"]) < len(panels["weekly"]))
def check_filing_corpus(r: Report) -> None:
section("13. Filing Corpus (5 checks)", r)
p = ROOT / "benchmark" / "daily" / "filing_corpus.parquet"
r.add("13.1 filing corpus parquet exists", p.exists())
if not p.exists():
return
fc = pd.read_parquet(p)
r.add("13.2 filing count >= 280K", len(fc) >= 280_000, f"got {len(fc)}")
r.add("13.3 per-ticker coverage >= 4000", fc["ticker"].nunique() >= 4000 if "ticker" in fc.columns else False,
f"got {fc['ticker'].nunique() if 'ticker' in fc.columns else 'no ticker col'}")
r.add("13.4 filing_type col present", "filing_type" in fc.columns)
r.add("13.5 path col present", "path" in fc.columns or "filing_path" in fc.columns, warn=True)
def check_xbrl_ontology(r: Report) -> None:
section("14. XBRL Ontology (3 checks)", r)
onto = ROOT / "xbrl" / "ontology" / "industry_ontology.json"
cat = ROOT / "xbrl" / "ontology" / "tag_catalog.parquet"
r.add("14.1 industry_ontology.json exists", onto.exists())
r.add("14.2 tag_catalog.parquet exists", cat.exists())
if onto.exists():
with open(onto) as f:
o = json.load(f)
# Ontology JSON is {by_sector: {...}, by_industry: {...}}; count sub-industries.
n_ind = len(o.get("by_industry", {})) if isinstance(o, dict) else 0
n_sec = len(o.get("by_sector", {})) if isinstance(o, dict) else 0
r.add("14.3 ontology has >= 5 sectors", n_sec >= 5, f"got {n_sec} sectors, {n_ind} industries")
def check_config_reproducibility(r: Report) -> None:
section("15. Config / Reproducibility (5 checks)", r)
r.add("15.1 START_DATE hardcoded", getattr(config, "START_DATE", None) == "2021-01-01",
f"got {getattr(config, 'START_DATE', None)}")
end = getattr(config, "END_DATE", None)
r.add("15.2 END_DATE hardcoded (not dynamic today)", end in ("2026-03-01", "2026-04-01"),
f"got {end}", warn=end not in ("2026-03-01", "2026-04-01"))
r.add("15.3 TEMPORAL_SPLIT_RATIO defined", hasattr(config, "TEMPORAL_SPLIT_RATIO"))
r.add("15.4 VALUATION_HOLDOUT_RATIO defined", hasattr(config, "VALUATION_HOLDOUT_RATIO"))
seed = getattr(config, "BENCHMARK_SEED", None)
r.add("15.5 BENCHMARK_SEED defined", seed is not None, f"seed={seed}", warn=seed is None)
def check_raw_sources(r: Report) -> None:
section("16. Raw Source Coverage (6 checks)", r)
r.add("16.1 prices/ non-empty", any((ROOT / "prices").glob("*")) if (ROOT / "prices").exists() else False)
r.add("16.2 fundamentals/ non-empty", any((ROOT / "fundamentals").glob("*")) if (ROOT / "fundamentals").exists() else False)
r.add("16.3 filings/ ticker count >= 4000",
len(list((ROOT / "filings").glob("*"))) >= 4000 if (ROOT / "filings").exists() else False)
r.add("16.4 xbrl/parsed non-empty",
any((ROOT / "xbrl" / "parsed").glob("*")) if (ROOT / "xbrl" / "parsed").exists() else False)
r.add("16.5 macro/ non-empty", any((ROOT / "macro").glob("*")) if (ROOT / "macro").exists() else False)
r.add("16.6 real_estate/ non-empty",
any((ROOT / "real_estate").glob("*")) if (ROOT / "real_estate").exists() else False)
def check_labels(r: Report, df: pd.DataFrame, gran: str) -> None:
section(f"17.{gran} Labels ({gran}, 3 checks)", r)
if "label" not in df.columns:
r.add(f"17.{gran}.1 label col present", False)
return
r.add(f"17.{gran}.1 label col present", True)
labels = set(df["label"].dropna().unique())
expected = {"lower_end_r2k", "small_cap_outside", "other"}
unexpected = labels - expected
r.add(f"17.{gran}.2 labels in expected set", len(unexpected) == 0, f"unexpected: {unexpected}")
r.add(f"17.{gran}.3 each label non-empty",
all(df["label"].value_counts().get(l, 0) > 0 for l in expected),
f"counts: {df['label'].value_counts().to_dict()}", warn=True)
def check_news(r: Report, gran: str) -> None:
"""News features land in benchmark/panel_train.parquet (via enrich_benchmark)."""
section(f"18.{gran} News Features (benchmark {gran}, 3 checks)", r)
p = ROOT / "benchmark" / gran / "panel_train.parquet"
if not p.exists():
r.add(f"18.{gran}.1 benchmark panel exists", False, str(p))
return
bp = pd.read_parquet(p, columns=None)
cols = bp.columns
r.add(f"18.{gran}.1 news_count_7d present", "news_count_7d" in cols)
r.add(f"18.{gran}.2 has_press_release_7d present", "has_press_release_7d" in cols)
if "news_count_7d" in cols:
nc = bp["news_count_7d"].dropna()
r.add(f"18.{gran}.3 news_count_7d >= 0", (nc >= 0).all(), f"min={nc.min()}")
def check_task_definition(r: Report, gran: str) -> None:
section(f"19.{gran} Task Definition JSON ({gran}, 3 checks)", r)
p = ROOT / "benchmark" / gran / "task_definition.json"
r.add(f"19.{gran}.1 task_definition.json exists", p.exists())
if not p.exists():
return
with open(p) as f:
td = json.load(f)
r.add(f"19.{gran}.2 task_definition is non-empty", bool(td))
expected_keys = {"granularity", "targets", "horizons", "evaluation"}
r.add(f"19.{gran}.3 has core benchmark keys", expected_keys.issubset(set(td.keys()) if isinstance(td, dict) else set()),
f"missing: {expected_keys - set(td.keys() if isinstance(td, dict) else [])}")
def check_valuation_tasks_json(r: Report) -> None:
section("20. Valuation Tasks JSON (3 checks)", r)
p = ROOT / "benchmark" / "daily" / "valuation_tasks.json"
r.add("20.1 valuation_tasks.json exists", p.exists(), warn=not p.exists())
if not p.exists():
return
with open(p) as f:
vt = json.load(f)
tasks = vt.get("tasks", []) if isinstance(vt, dict) else []
r.add("20.2 >= 6 valuation tasks defined (A-F)", len(tasks) >= 6, f"got {len(tasks)}")
def _has_id(t):
if isinstance(t, str):
return len(t) > 0 # task IDs are strings like "A_valuation_accuracy"
if isinstance(t, dict):
return any(k in t for k in ("name", "id", "task_id", "task"))
return False
r.add("20.3 each task has an identifier", all(_has_id(t) for t in tasks) if tasks else False)
def main() -> int:
r = Report()
r.lines.append("MacroLens Expanded Validation Suite")
r.lines.append("=" * 50)
check_universe(r)
check_config_reproducibility(r)
check_raw_sources(r)
check_xbrl_ontology(r)
check_cross_granularity(r)
check_filing_corpus(r)
check_valuation_tasks_json(r)
for gran in ("daily", "weekly", "monthly"):
df = check_panel_schema(r, gran)
if df is None:
continue
check_prices(r, df, gran)
check_xbrl_balance(r, df, gran)
check_derived(r, df, gran)
check_macro(r, df, gran)
check_filing_context(r, df, gran)
check_labels(r, df, gran)
check_news(r, gran)
check_split_integrity(r, gran)
check_scenarios(r, gran)
check_benchmark_files(r, gran)
check_valuation_holdout(r, gran)
check_task_definition(r, gran)
r.lines.append("")
r.lines.append("=" * 50)
r.lines.append(f"TOTAL: {r.total} checks | PASS: {r.pass_count} WARN: {r.warn_count} FAIL: {r.fail_count}")
print("\n".join(r.lines))
return 0 if r.fail_count == 0 else 1
if __name__ == "__main__":
sys.exit(main())