Free lesson · Point-in-time data

Reproducible snapshots and discrepancy ledgers

Open interactive lessonPractice calculationsExplore labs

A research result needs a record of exactly which inputs and transformations produced it.

A manifest should include source URL or dataset identifier, retrieval time, permitted use, timezone, units, universe definition, version, transformation settings and a content checksum. Record code revision and model configuration alongside the data.

A checksum detects byte differences, not whether a dataset is true, timely or licensed. When reproductions differ, categorize the discrepancy: source revision, units, sample membership, code change, randomness or numerical tolerance. Never replace an unexplained discrepancy with an edited target number.

Apply it in a strategy

• Freeze inputs at the stated decision time and record their units.
• Hand off research so another person can reproduce both its data and calculations.
• Recompute the example, then change the material assumption and explain the difference.

Research deliverable

Reproducible snapshots and discrepancy ledgers: produce the worked calculation, a timestamped input record and a written decision addressing this limitation: Matching a published number can coexist with shared data leakage or a common implementation error.

These are synthetic mechanics examples, not historical performance or paper replications. Module evidence and research boundaries record the 12 September 2026 review.

Evidence and boundaries · reviewed 12 September 2026

The records below distinguish research status, access depth and data dates. Abstract-only review identifies research questions; it does not establish a replicated empirical claim.

Further reading: Crane, Karra & Soto: Total Recall? Evaluating the Macroeconomic Knowledge of Large Language Models ↗

Federal Reserve Finance and Economics Discussion Series working paper 2025-044. Version: June 2025; later versions not established. Review: 2026-09-12; Abstract on Federal Reserve 2025 research index only. Markets: Macroeconomic data and release dates. Data dates: Exact series, date ranges and model snapshots not inspected. Limitation: Motivates explicit vintage and availability checks. The lesson does not quantify model error rates or reproduce the study.

Further reading: Federal Reserve Bank of St. Louis: FRED API real-time periods ↗

Living primary provider documentation. Version: Reviewed 12 September 2026. Review: 2026-09-12; Documentation page. Markets: Series-specific macroeconomic data. Data dates: Series-dependent; no dataset downloaded. Limitation: API semantics do not establish the exact intraday availability of every vintage; source release timestamps and time zones must be verified separately.

Further reading: SEC: EDGAR Application Programming Interfaces ↗

Living primary provider documentation. Version: Reviewed 12 September 2026. Review: 2026-09-12; Documentation page. Markets: US public-company submissions and XBRL facts. Data dates: Filing-specific; no archive downloaded. Limitation: Facts can have amendments, duplicate contexts and differing units. An annual fiscal period is not an information-release time.

Research sources, review dates and limitations

Connect the ideas: Information and decision time

Retrieve: Use only information available when the decision is made.

Check the change: Observation dates, release delays, revisions and label maturity require different availability checks.

Statistics → Research & backtests → Time series → Research & robust tuning → Financial machine learning → Putting it all together

# Python 3.10+; standard library and NumPy only.
# Synthetic teaching inputs; conventions and units are defined in the notation above.
import hashlib
import json

def snapshot_check(records,reproduced,reference):
    payload=json.dumps(records,sort_keys=True,separators=(',',':'),allow_nan=False).encode('utf-8')
    return hashlib.sha256(payload).hexdigest(),abs(reproduced-reference)

checksum,error=snapshot_check([{'time':1,'value':10}],.0118,.012)
assert len(checksum)==64 and abs(error-.0002)<1e-12
print(checksum,error)