DATASHEET.md

# Datasheet for MacroLens

This datasheet follows the *Datasheets for Datasets* framework (Gebru et al., *Communications of the ACM*, 2021).

---

## 1. Motivation

**For what purpose was the dataset created?**
MacroLens evaluates forecasting and valuation models that must reason over numerical history *and* contextual information — macroeconomic state, scenarios, and firm text — in a financial setting. It addresses gaps in three existing benchmark families: generic time-series forecasting benchmarks drop text and valuation tasks; financial language benchmarks drop forecasting and event reasoning; recent context-rich forecasting datasets are non-financial or omit valuation.

**Who created the dataset and on behalf of which entity?**
Anonymous (NeurIPS 2026 Datasets & Benchmarks Track double-blind submission). Authors and affiliations to be disclosed after author notification.

**Who funded the creation of the dataset?**
Anonymous (will be disclosed after notification).

**Any other comments?**
The benchmark targets the *intersection* of contextual time-series forecasting, valuation, and scenario-conditioned event prediction, which prior public benchmarks have not covered jointly.

---

## 2. Composition

**What do the instances that comprise the dataset represent?**
A MacroLens instance is the tuple ⟨ticker $i$, timestamp $t$, granularity $g$, lookback panel $x_{i,t-L:t,g}$, static covariates $z_i$, optional scenario $s_t$, optional text $u_{i,\le t}$⟩, paired with a task-specific target $y_{i,t}$.

**How many instances are there in total?**
- 4,841,094 daily panel rows (3,219,018 train / 1,622,076 test) over 4,416 tickers and 1,313 trading days.
- 1,009,314 weekly panel rows.
- 232,483 monthly panel rows.
- 23,147 T2 valuation ground truths.
- 23,147 T5 private-valuation ground truths (same 1,324 holdout tickers, price-stripped).
- ~14,500 T3 (ticker, fiscal year, field) ground truth tuples (11-field curated dense panel).
- 4,072,843 T4 scenario-forecast ground-truth rows from 1,622,076 test panel rows × 1,130 events.
- 11,065 T6 generator-evaluation ground truths.
- 23,367 T7 real-estate ground truth rows over 23,190 unique addresses.
- 1,130 macroeconomic scenario events across 49 types.

**Does the dataset contain all possible instances or is it a sample (e.g., a sample of a larger set)?**
The 4,416-ticker universe is the union of: full Russell 2000 (1,923 IWM holdings), full S&P SmallCap 600 (72 IJR-only additions), iShares Micro-Cap (225 IWC additions), and the 2,196 small-cap NASDAQ/NYSE tickers outside all three indices, filtered to company market cap ≤ \$7.4B. This is **not** a sample — it is the complete enumeration of U.S. small/micro-cap equities meeting the universe spec on the trade dates 2021-01-04 through 2026-03-31.

**What data does each instance consist of?**
- **Numeric panel** (131 features per (ticker, date) coordinate): 6 OHLCV + 19 derived valuation ratios + 45 XBRL statement fields with TTM rolling-sum variants + 46 FRED macro + 7 EIA commodity + 1 days-since-filing + 7 index/membership flags.
- **Static covariates**: ticker metadata (sector, industry, exchange), security_type (operating / fund / SPAC), index memberships.
- **Scenario object** (optional, T4): event_type (49 categories), structured natural-language description, scenario_id.
- **Text** (optional): SEC filings (markdown + PDF), financial news articles.
- **Target**: per-task, see Section "Splits" below.

**Is there a label or target associated with each instance?**
Yes, per task (T1: horizon-length close trajectory; T2/T5: realized market cap; T3/T6: 11 canonical XBRL field values; T4: 63-day post-event return percentage; T7: rent + price).

**Is any information missing from individual instances?**
Yes, by point-in-time design. Quarterly XBRL facts apply a post-acceptance lag (so they appear in $x_{i,t,g}$ only after the publication timestamp). News articles enter only after publication. The 14 tickers without XBRL or yfinance fundamentals (5 FDIC-only banks + 9 SEC-empty stubs that yfinance also fails) are applicability-masked on T2/T3/T5/T6 (kept for T1, T4 with prices+filings only).

**Are relationships between individual instances made explicit?**
Yes. Tickers are linked to scenarios via dates and event_id. Real-estate addresses link to metros. Filings link to tickers via CIK. All keys are stored as identifier columns, not derived joins.

**Are there recommended data splits?**
- **T1, T4 (forecasting)**: chronological 70/30 split at **2024-09-03** (1,622,076 daily test rows).
- **T2, T3, T5, T6 (valuation + generation)**: 30% company-level holdout = **1,324 tickers, seed = 42**. Each ticker contributes its latest valid snapshot. T3, T6 add a per-ticker temporal split (latest fiscal year for test, prior years for train).
- **T7 (real-estate)**: 30% address-level holdout (random, seeded).

**Are there any errors, sources of noise, or redundancies in the dataset?**
- yfinance occasionally yields stale or misaligned quarter-close fundamentals; the loader applies a one-day lag for safety.
- The 14 fundamentals-empty tickers are applicability-masked, not excluded.
- T4 events with pre-event price below SEC penny-stock threshold ($0.50) are dropped at build time (~140 rows) because percentage-return arithmetic blows up at the noise floor.
- T7 has 854 duplicate-address rows in the train pool (53,804 unique vs 54,658 raw); deduplicated at canonical-index time.

**Is the dataset self-contained, or does it link to or otherwise rely on external resources?**
The Hugging Face release is bundled-self-contained for SEC EDGAR (filings + XBRL facts), FRED + EIA macro series, yfinance-derived prices + fundamentals, and the curated benchmark parquets — no user credentials needed for these. Two sources are **gated by external licensing** and ship as derived features + reconstruction scripts only:

| Source | Bundled in HF release? | User credentials required for raw re-fetch? |
|---|---|---|
| SEC EDGAR (filings, XBRL) | Yes (public domain) | No (free) |
| FRED, EIA (macro) | Yes (public domain) | No (free; FRED API key recommended for high rate) |
| yfinance (prices, fundamentals) | Yes (derived features) | No (free) |
| Macroeconomic event scenarios | Yes (curated by us, CC-BY-4.0) | No |
| **RentCast** (real estate raw) | **NO — derived features only** | **YES — user's own RentCast subscription** for `collect_real_estate.py` |
| **Financial news** (~215k articles) | **NO — derived counts only** | **YES — user's own news-API key** for `collect_news.py` |

**Does the dataset contain data that might be considered confidential?**
No. All sources are public regulatory filings (SEC EDGAR), public market data (yfinance), public macroeconomic series (FRED, EIA), and licensed real-estate listings (RentCast, used under their terms).

**Does the dataset contain data that, if viewed directly, might be offensive, insulting, threatening, or might otherwise cause anxiety?**
No, beyond standard financial-news content (corporate disputes, lawsuits, layoffs) which is part of public regulatory disclosure.

**Does the dataset relate to people?**
Indirectly — SEC filings name corporate officers and directors as part of public regulatory disclosure (the same information that appears on EDGAR). No private individuals; no PII beyond what is in public regulatory filings.

**Does the dataset identify any subpopulations?**
The dataset records `security_type` (operating, fund, SPAC) and Global Industry Classification Standard (GICS) sector for every ticker. No protected demographic categories.

---

## 3. Collection Process

**How was the data associated with each instance acquired?**
- **Universe**: iShares IWM/IJR/IWC ETF holdings + NASDAQ Trader symbol directory (filtered to market cap ≤ \$7.4B).
- **Prices**: Yahoo Finance.
- **Fundamentals**: yfinance (3.22M rows) + SEC EDGAR XBRL company-facts API (46.79M facts, 92.6% coverage).
- **Macro**: FRED + EIA via the publicly documented APIs.
- **Filings**: SEC EDGAR (10-K, 10-Q, 8-K, 20-F, 6-K, N-CSR, N-CSRS).
- **News**: provider feed + entity linking.
- **Real estate**: RentCast API (100 U.S. metros, 139,855 properties × 544 RentCast variants).

**What mechanisms or procedures were used to collect the data?**
Custom Python scripts (`collect_*.py`) using each source's official documented API. Rate limits were honored. All scripts are included in the release.

**If the dataset is a sample from a larger set, what was the sampling strategy?**
Not a sample — full enumeration of the universe spec over 2021-01-04 → 2026-03-31. Within that, the 30% company-level valuation holdout uses **stratified sampling** on (sector, market-cap quartile) at fixed seed = 42.

**Who was involved in the data collection process?**
Anonymous authors. No human annotators (the dataset uses programmatic API queries).

**Over what timeframe was the data collected?**
Source data was published over 2021-01-04 — 2026-03-31. Collection scripts were run in 2025-2026 to assemble the panel.

**Were any ethical review processes conducted?**
N/A — public-records data only.

---

## 4. Preprocessing / Cleaning / Labeling

**Was any preprocessing/cleaning/labeling of the data done?**
Yes:
- **Point-in-time alignment**: every observation aligns to publication timestamp (filings post-acceptance lag, quarterly XBRL post-acceptance, news post-publication).
- **Algebraic-leakage scrubbing for T2/T5**: every input column is auto-tested against $\log y$; any column with $|\text{Pearson}| > 0.99$ to the target is excluded. Largest residual T2 correlation post-scrub is shares-outstanding at $\rho = 0.30$, a legitimate size proxy.
- **APE clipping at 10×** (1,000%) on all valuation tasks to prevent a single mispredicted outlier from dominating MAPE-style metrics.
- **Outlier cleanup at source for T4**: rows with pre-event price below SEC penny-stock threshold (\$0.50) dropped at build time.
- **Address deduplication for T7** at canonical-index time (854 duplicates in train pool, 177 in eval pool removed).
- **TTM rolling-sum variants** computed for flow-style XBRL fields (revenue, net income, etc.).

**Was the "raw" data saved in addition to the preprocessed/cleaned/labeled data?**
Yes. The release bundles raw XBRL facts (`xbrl/`), raw prices (`prices/`), raw fundamentals (`fundamentals/`) alongside the curated `benchmark/` parquets so downstream researchers can re-derive features.

**Is the software that was used to preprocess/clean/label the data available?**
Yes — `preprocess.py`, `assemble_benchmark.py`, `build_ontology.py`, `enrich_benchmark.py`, `generate_scenarios.py`, `build_valuation_tasks.py`, `validate_all.py`. All under MIT.

---

## 5. Uses

**Has the dataset been used for any tasks already?**
Yes — the accompanying paper reports a 17-method baseline panel across 7 families on T1–T7.

**Is there a repository that links to any or all papers or systems that use the dataset?**
The HF dataset card (this README) will track citations. Currently: the accompanying NeurIPS 2026 D&B paper.

**What (other) tasks could the dataset be used for?**
- Multi-modal time-series forecasting research.
- Macroeconomic-event impact studies.
- LLM evaluation under domain-specific (financial) tasks.
- Private-market valuation modeling.
- Cross-domain transfer (real-estate vs equity valuation).
- Scenario reasoning + counterfactual forecasting.

**Is there anything about the composition of the dataset or the way it was collected that might impact future uses?**
- U.S.-only and English-only — international generalizability not supported.
- Survivorship bias is partially mitigated by including delisted tickers, but pre-2021 history is not covered.
- The 30% company-level holdout for T2/T3/T5/T6 evaluates OOD-ticker generalization but not OOD-sector or OOD-industry by construction.

**Are there tasks for which the dataset should not be used?**
- **Trading decisions**: the dataset is a research benchmark; metrics do not include transaction costs, slippage, or execution modeling. Direct trading use is **not recommended**.
- **International generalizability claims**: U.S. equities only.
- **Deployment safety**: no adversarial-robustness testing.

---

## 6. Distribution

**Will the dataset be distributed to third parties outside of the entity on behalf of which the dataset was created?**
Yes — Hugging Face Datasets, public.

**How will the dataset be distributed?**
- Primary: `huggingface.co/datasets/macrolens/MacroLens` (Croissant-validated, NeurIPS-D&B compliant).
- Code: same repo.
- Reconstruction scripts: same repo (raw filings + news re-fetched from official sources).

**When will the dataset be distributed?**
Public at time of NeurIPS 2026 D&B-track submission.

**Will the dataset be distributed under a copyright or other intellectual property license, and/or under applicable terms of use (ToU)?**
- Derived features + curated panel: **CC-BY-4.0**.
- Code: **MIT**.
- Vendored libraries (under `methods/_vendored/`): TSLib (MIT), ModernTCN (Apache 2.0).
- Reconstruction scripts: MIT.

**Have any third parties imposed IP-based or other restrictions on the data associated with the instances?**
- yfinance, FRED, EIA, RentCast: each has its own ToU; the release ships derived features and reconstruction scripts.
- SEC EDGAR: public domain.

**Do any export controls or other regulatory restrictions apply to the dataset or to individual instances?**
No.

---

## 7. Maintenance

**Who is supporting/hosting/maintaining the dataset?**
Anonymous (NeurIPS 2026 D&B submission). Maintainer-of-record will be disclosed after author notification.

**How can the owner / curator / manager of the dataset be contacted?**
Through the Hugging Face dataset discussions tab (`huggingface.co/datasets/macrolens/MacroLens/discussions`) or via the corresponding-author email (post-notification).

**Is there an erratum?**
None at submission. Errata will be tracked in the dataset card's `Changelog` section.

**Will the dataset be updated?**
Yes — minor versioned updates planned to extend the time window and refresh upstream sources. Versioning follows semver; each release tags a Git-style snapshot in the HF repo.

**If the dataset relates to people, are there applicable limits on the retention of the data associated with the instances?**
N/A — public regulatory filings only.

**Will older versions of the dataset continue to be supported/hosted/maintained?**
Yes — older revisions remain accessible via HF dataset revision tags (commit SHAs).

**If others want to extend/augment/build on/contribute to the dataset, is there a mechanism for them to do so?**
Yes — pull requests via the HF dataset repo or the GitHub mirror. Contributions are reviewed for license compatibility (CC-BY-4.0 compatible only) and benchmark protocol consistency.

---

*This datasheet was prepared at the time of NeurIPS 2026 D&B-track submission. The Croissant metadata file (auto-generated by Hugging Face) at `https://huggingface.co/api/datasets/macrolens/MacroLens/croissant` is the machine-readable counterpart to this human-readable datasheet.*