GLOSTAT — Evidence-based Probability Predictor for Global Equities

Calibrated multi-horizon predictions with evidence chains.

Information tool. Not investment advice. Past calibration ≠ future performance.

Status

v2.2.0 ACTIVE (2026-09-10) — ThesisRegistry, measurement gate, three new thesis slots (52w high, KR short-interest anomaly, US→KR spillover). New theses ship at n=0 / weight=0 until a hindcast measures them. Failures stay on the calibration table. All v1.x releases are yanked from PyPI; pip install glostat now resolves to v2.2.0+.

v2.0.0 removed the sizing-tier attachment (INV-GS-111 deprecated); v2.0.1 hardened API-key handling in DART/ECOS/KIS clients.

v0.6/v0.7 framed GLOSTAT as a decision engine (BUY/SELL action output) and concluded "8 thesis FAIL" against a Sharpe ≥ 0.8 / AUC ≥ 0.62 / OOS deg ≤ 30% gate. v1.0 reframed the same data as a prediction tool: probability distribution + 1-sigma (~68%) CI per INV-GS-113 + per-thesis Brier-weighted contribution + evidence chain. The 8-thesis FAIL outcomes are now the first input rows of the calibration table — E_PEAD AUC 0.587, E_FOREIGN_REVERSAL OOS Sharpe 1.46, E_FOMC_DRIFT AUC 0.357 (anti-predictor) — all carry honest, sample-aware weights.

v2.0 migration: Prediction.dca_sizing field is REMOVED. The predictor surface is strictly probability + CI. Callers that read .dca_sizing must remove that access. See v2.0 release notes.

Read first: docs/post_mortem/SPRINT5_FAIL_post_mortem.md (the v0.6 honest diagnosis), then docs/ssot/PLAN_v1.0.md (the active spec), then docs/CALIBRATION.md (the empirical predictive strength table), then docs/ROADMAP_v2.md (v2.x.x integrated roadmap).

What this project IS

• A calibrated probability predictor — outputs Prediction(up_probability, confidence_interval_bps, contributing_signals, ...) with Brier-derived ensemble weights per thesis.
• A deterministic hindcast harness — turns any thesis into a calibration row (Brier + AUC + Sharpe + OOS) with explicit IS/OOS split and reproducibility guarantees.
• A snapshot broker — every external data response persisted as a parquet shard + SQLite index + Merkle leaf, so any prediction can be replayed bit-for-bit months later.
• An open-source research framework — MIT, fork-friendly, designed so third-party thesis authors can plug in and contribute calibration data.
• A compliance gate that makes broadcast permanently impossible (ComplianceError on Telegram / mass-email entry points; INV-GS-024) and stamps a personal-use, not-investment-advice disclaimer on every prediction (INV-GS-104).
• A prompt registry that pins each LLM call to a sha256 so the prompt graph is auditable across versions.
• 71+ numbered invariants (INV-GS-001..114) with a 1:1 unit-test mapping and a machine-readable configs/invariants.yaml. Active invariants include INV-GS-101 (no BUY/SELL output), INV-GS-104 (per-prediction disclaimer), INV-GS-113 (CI 1-sigma + p-value + all-noise statistical disclosures), INV-GS-114 (KR megacap honesty footer).

What this project IS NOT

• Not investment advice. Use at your own risk. Read the post-mortem first.
• Not a trading bot. No BUY/SELL action output (INV-GS-101). No target/stop/size output. The Prediction dataclass deliberately omits any field that prescribes action.
• Not an alpha-generating decision engine. v0.6 attempted that, failed honestly across 8 thesis. The framework's value is the **honest measurement
  • Brier weighting**, not a guaranteed alpha.
• Not a black-box predictor. Every Prediction carries a SignalContribution chain with calibration window, n_samples, calibration_auc, confidence_v2, and source snapshot IDs.
• Not a broadcast tool. broadcast_telegram and mass_email are inert sentinels that always raise (INV-GS-024).
• Not a multi-user product. Personal use only.

Supported markets

Active data clients (13): yfinance, sec_edgar, cftc, ccxt, naver_kr, dart (KR insider/disclosure), ecos (BoK macro), kis (KIS Open API read-only), krx_short (KRX short statistics), toss (Toss local-parquet cache), commodity, kr_calendar, bigdata (phase-gated, MVP-blocked per INV-GS-036).

KR predictions use up to 11 active KR-specific signals: E_FUNDAMENTAL_KR (yfinance PER/ROE/dividend + DART overlay), E_FUNDAMENTAL_KR_CYCLICAL, E_FOREIGN_REVERSAL (3-source Naver+KIS+Toss flow fusion), E_INSIDER_KR (DART elestock), E_MACRO_KR (BoK ECOS), E_PEAD_KR, E_INSIDER_VELOCITY_KR, E_SHORT_SELLING_KR, E_INTRADAY_FLOW_KR, E_COMMODITY_INDEX_KR, E_TIME (Ichimoku, universe-agnostic). See docs/KR_SUPPORT.md for the full guide.

# Live KR prediction (no Bigdata MCP, $0 cost, no SEC_USER_AGENT needed for KR)
glostat predict 005930 --horizon swing_5d   # 삼성전자
glostat predict 096770 --horizon swing_5d   # SK Innovation

# Live US prediction (SEC EDGAR requires User-Agent per INV-GS-038)
GLOSTAT_SEC_USER_AGENT="Your Name your.email@yourdomain.com" \
  glostat predict AAPL --horizon swing_5d

What we tested (8 thesis → calibration data, not failures)

The numbers below come from the v0.6/v0.7 hindcast runs preserved in cache/hindcast/ and reframed here as the v1.0 calibration baseline.

*Brier-derived weight (illustrative — actual values computed at run time). Full table + interpretation: docs/CALIBRATION.md.

The v0.6 verdict on the same data: "8 thesis FAIL → automatic shutdown." The v1.0 verdict: "8 calibrated signals, composite up_probability exists with explicit confidence interval, weak/anti-predictive signals carry near-zero weight."

Both readings are honest. v1.0+ is the more useful one.

The above table is the v0.6 baseline only. Current main has 21 thesis modules across US + KR + crypto, including the v1.x additions E_FUNDAMENTAL_KR, E_PEAD_KR, E_INSIDER_VELOCITY_KR, E_ANALYST_REVISION, E_SHORT_SELLING_KR, E_INTRADAY_FLOW_KR, E_FUNDAMENTAL_KR_CYCLICAL, E_COMMODITY_INDEX_KR, E_MACRO_KR, E_INSIDER_KR. See docs/CALIBRATION.md for the full per-thesis table and docs/KR_SUPPORT.md for KR-specific signal documentation.

KR megacap honesty footer (INV-GS-114)

Phase KR M1 hindcast on KOSPI 200 (n = 3,510 samples) measured AUC ≤ 0.51 — at the edge of statistical noise. The CLI surfaces this footer on every KR megacap prediction (*** KR megacap universe — AUC ≤ 0.51, predictions are weak signals). This is the same honest data that drives the v1.x E_FOREIGN_REVERSAL Brier weight collapse on that universe; the framework shows it rather than hides it.

Architecture overview

                ┌─────────────────────────────────────────────────────┐
                │                CLI / library entry                   │
                │   glostat predict <ticker>      glostat calibrate    │
                └─────────────────────────────────────────────────────┘
                                       │
                ┌──────────────────────┼──────────────────────┐
                ▼                      ▼                      ▼
        DataRouter             Compliance Gate         PromptRegistry
        (phase-gated)          (broadcast=ERROR)       (sha256 per call)
                │
   ┌────────────┼────────────┬────────────┐
   ▼            ▼            ▼            ▼
yfinance    SEC EDGAR    CFTC/CCXT    Bigdata MCP   ← Phase 2+, blocked in MVP
   │            │            │            │
   └────────────┴────────────┴────────────┘
                │
                ▼
        Snapshot Broker  ───►  parquet shards + SQLite index + Merkle leaves
                │
                ▼
            Thesis modules  ───►  raw_score, direction, sources
                │
                ▼
   ┌─────────────────────────────────────┐
   │  predictor/composite.py             │
   │   predict()  +  Brier weights       │   ← INV-GS-103
   └─────────────────────────────────────┘
                │
                ▼
            Prediction  ───►  up_probability + 1-sigma CI + SignalContribution[] + disclaimer
                │
                ▼  (calibration loop, quarterly)
       calibration_table.parquet
                │
                ▼
        docs/CALIBRATION.md (auto-regenerated)

Snapshot broker

from datetime import UTC, datetime
from pathlib import Path

from glostat.data.snapshot_broker import SnapshotBroker, SnapshotKey

broker = SnapshotBroker(root=Path("./snapshots"))
key = SnapshotKey(
    uaid="XNAS.AAPL",
    edge_type="tearsheet",
    ts_utc=datetime(2026, 4, 28, 12, 0, tzinfo=UTC),
    tool="yfinance.fundamentals",
    params_canon='{"period":"quarter"}',
)
record = broker.save_snapshot(key, {"per": 28.4, "roe": 0.21})
print(record.leaf.leaf_hash[:12], broker.audit_root()[:12])
broker.close()

Calibrated prediction (v2.0)

from glostat.predictor import predict, load_calibration
from glostat.predictor.types import Prediction

cal_table = load_calibration()
prediction: Prediction = predict(
    ticker="AAPL",
    horizon="swing_5d",                     # intraday | swing_5d | swing_30d | long_3y
    contributions=(...),                    # build via collect_contributions(); see docs/EXAMPLES.md
    cal_table=cal_table,
)

print(f"p_up = {prediction.up_probability:.3f}")
low_bps, high_bps = prediction.confidence_interval_bps
print(f"  CI 1-sigma (~68%) bps = [{low_bps:+.1f}, {high_bps:+.1f}]")
if low_bps <= 0 <= high_bps:
    print("  *** includes 0 — no clear direction (INV-GS-113 X2)")

for c in prediction.contributing_signals:
    if c.direction == "skip":
        continue
    print(f"  {c.name:24} dir={c.direction:4}  "
          f"AUC={c.calibration_auc:.3f}  n={c.n_samples}")

print(prediction.disclaimer)                 # always non-empty (INV-GS-104)

Compliance gate (cannot be bypassed)

from glostat.risk.compliance_gate import broadcast_telegram, ComplianceContext

broadcast_telegram(
    ctx=ComplianceContext(user_profile_hash="0" * 64, jurisdiction="US"),
    chat_ids=["@anyone"], message="anything",
)
# → glostat.risk.compliance_gate.ComplianceError: INV-GS-024 …

Quickstart

Requires Python ≥ 3.11.

# install (PyPI default = v2.0.1+, since v1.x are yanked)
pip install glostat

# or from source (uv preferred)
git clone https://github.com/deximple/glostat.git
cd glostat
uv sync --extra dev

# verify
uv run pytest -q                                   # 836+ test functions
uv run python -c "import glostat; print(glostat.__version__)"   # → 2.0.1

# Mock prediction (no network, fixture data)
uv run glostat predict AAPL --horizon swing_5d --mock

# JSON output (machine-readable)
uv run glostat predict AAPL --horizon swing_5d --mock --json

# Refresh calibration_table.parquet from cached hindcast reports
uv run glostat calibrate --out cache/calibration_table.parquet

# Run KR hindcast (KOSPI 200 universe, produces calibration JSON)
uv run glostat kr-hindcast --universe kospi200 --start 2024-01-01 --end 2026-04-30

US live mode requires GLOSTAT_SEC_USER_AGENT="Your Name your.email@yourdomain.com" (SEC EDGAR mandates a contactable User-Agent — INV-GS-038). KR live mode does NOT need it (Naver / DART / ECOS / KIS use their own keys; bare glostat predict 005930 --horizon swing_5d works).

Reusable for your own thesis

The infrastructure is independent of which thesis you screen. To add a new thesis to the calibration table:

1. Write a thesis module. Subclass the Thesis protocol in src/glostat/experts/, return a typed (direction, raw_score, sources). See docs/EXAMPLES.md for a working template.
2. Register a data source if needed. Add a routing entry in src/glostat/data/data_router.py. The DataRouter enforces phase gating so paid sources stay blocked until you explicitly opt in.
3. Run the hindcast. Configure Hindcast, point at a universe, get an IS/OOS report with AUC, Sharpe, Brier.
4. Add a calibration row. Append the result to cache/calibration_table.parquet (one row per thesis-universe-horizon triple). The Brier-weighted ensemble picks the weight automatically.
5. PR with calibration data attached. New thesis PRs must include n ≥ 50, AUC, Sharpe, OOS deg and a calibration row. (INV-GS-026 + INV-GS-105.)

Full walkthrough: docs/EXAMPLES.md. Migration from v0.7: docs/MIGRATION_v0.7_TO_v1.0.md.

Repo layout

src/glostat/
  core/         # Verdict (deprecated, kept for back-compat), shared types/errors
  data/         # 13 data clients (snapshot broker, free-stack clients, phase-gated DataRouter)
  experts/      # 21 thesis modules (PEAD, FOREIGN_REVERSAL, INSIDER_CLUSTER, FX_CARRY,
                #   E_FUNDAMENTAL_KR, E_PEAD_KR, E_INSIDER_VELOCITY_KR, E_ANALYST_REVISION, …)
  predictor/    # composite.predict(), confidence_v2 (5-component geometric mean), calibration I/O
                #   Prediction lives here (predictor/types.py), NOT in core/
  gating/       # cost gate, regime gate (kept; used during hindcast as calibration mask)
  replay/       # hindcast harness, sprint4_gate (now calibration check), kill criteria
  risk/         # compliance gate (INV-GS-024 + INV-GS-104)

configs/
  invariants.yaml    # 71 numbered invariants (001..114), v0.6 + INV-GS-111 deprecated entries flagged
  budget.yaml        # phase-gated budget caps (mvp $0)
  markets.yaml       # XNAS + XNYS + XKRX + XKOS
  gating.yaml        # cost / regime / anti-herd parameters (decision-engine vintage; used as calibration mask only in v1.0+)
  kill_criteria.yaml # narrowed v1.0 triggers (compliance, broker integrity, stale calibration)
  universes/

cache/
  calibration_table.parquet  # quarterly-updated weights per thesis-universe-horizon
  hindcast/                  # phase1b + phase1c + phase1d + phase_kr historical reports

tests/                       # 836+ test functions across 79 files; INV-GS-001..114 coverage
docs/
  ssot/                      # plan history v0.1 → v0.7 + PLAN_v1.0.md (active spec)
  post_mortem/               # honest Sprint 5 FAIL diagnosis (v0.6)
  research/                  # design notes
  ROADMAP_v2.md              # v2.x.x integrated roadmap (10-agent synthesis)
  v2.1_PRD.md                # v2.1 product requirements
  CALIBRATION.md             # per-thesis empirical predictive strength
  CONFIDENCE_V2.md           # 5-component confidence model (INV-GS-112)
  MIGRATION_v0.7_TO_v1.0.md  # developer migration guide
  EXAMPLES.md                # extending the framework

Honest reading order

If you are evaluating whether to adopt or fork this:

1. docs/post_mortem/SPRINT5_FAIL_post_mortem.md — start here. The v0.6 framework worked; the alpha didn't. v1.0 turns that honest finding into the calibration baseline.
2. docs/ssot/PLAN_v1.0.md — active framework spec. Section 1 explains the reframe rationale; Section 2 explains how the 8 FAIL outcomes become calibration data; later sections list invariants INV-GS-101..114 and deprecated INV-GS-001/005/033/111.
3. docs/CALIBRATION.md — empirical predictive strength of every thesis currently in the calibration table.
4. docs/MIGRATION_v0.7_TO_v1.0.md — developer migration guide.
5. configs/invariants.yaml — the contract the framework enforces.
6. docs/EXAMPLES.md — practical extension recipes.

Third-party data sources

Usage of the Bigdata.com integration requires a valid API key and is subject to Bigdata.com's Terms of Service. Users must obtain their own credentials independently. The Bigdata MCP client is gated behind GLOSTAT_PHASE=phase_2 (INV-GS-036) and is not active in the default MVP configuration.

Compliance disclaimer

GLOSTAT v1.0 is an information tool for personal use. Output is a probability distribution with explicit confidence intervals and source provenance — not an investment recommendation, not a securities solicitation, not financial advice. Past calibration data does not guarantee future predictive performance. Users are responsible for their own decisions.

broadcast_telegram and mass_email raise ComplianceError permanently and unconditionally (INV-GS-024). Every Prediction instance carries a non-empty disclaimer field, validated at construction time (INV-GS-104).

Contributing

Issues and pull requests are welcome. Useful directions:

• New thesis modules that screen a different thesis (event-driven, cross-asset momentum, options-implied, factor-based). Must include calibration data.
• New data source clients (Polygon free tier, Tiingo, Stooq, FRED) routed through DataRouter with phase gating.
• Refinements to Brier weighting / sample-size guards in predictor/composite.py.
• Fixes / hardening of the snapshot broker, hindcast harness, or compliance gate.

PR template enforces:

• New thesis → calibration row in cache/calibration_table.parquet (INV-GS-026, INV-GS-105)
• No new INV-GS-101 violation (no BUY/SELL output)
• No new INV-GS-024/104 weakening (no broadcast, no missing disclaimer)

See CONTRIBUTING.md for coding style, INV-GS conventions, and templates.

License

MIT — see LICENSE. Use it commercially, fork it, embed it, port it, just keep the copyright notice.

Citing

If this framework helps your research or post-mortem write-up, a link back to this repository is appreciated. Cite the calibration table version (v2.x.x, quarter-bumped) so reproducibility is preserved.