MacroLens

A benchmarking corpus for contextual financial reasoning under macroeconomic scenarios across 4,416 U.S. small- and micro-cap equities (2021-01-04 — 2026-03-31). MacroLens unifies seven tasks over a single point-in-time panel: contextual time-series forecasting, public valuation, financial-statement generation, scenario-conditioned return forecasting, private-company valuation, generator evaluation from natural-language descriptions, and real-estate valuation.

Task	Type	Output
T1 Contextual Forecasting	Time-series	Horizon-length close trajectory
T2 Public Valuation	Tabular regression	Equity market cap
T3 Financial Statement Generation	Structured generation	11 canonical XBRL fields per (ticker, fiscal year)
T4 Scenario-Conditioned Return	Event forecasting	63-day post-event return percentage
T5 Private-Company Valuation	Tabular regression (price-stripped)	Equity value w/o market data
T6 Generator Evaluation	NL→ structured	Same 11 fields from a natural-language company description
T7 Real-Estate Valuation	Cross-domain regression	Rent + price per RentCast address

Every instance carries a 131-numeric / 141-column point-in-time panel (prices, 46.8M XBRL accounting facts, 53 macroeconomic series, filing recency, derived ratios), an optional macroeconomic scenario object (1,130 events across 49 types), and optional SEC filings + financial-news context. Temporal alignment is strictly point-in-time: every observation visible at prediction timestamp $t$ was publicly available by $t$.

Quickstart

import macrolens as ml

# 1. Load (X, y, meta) — identical schema across train/test
X_train, y_train, meta_train = ml.load("T1", "train", granularity="daily")
X_test,  y_test,  meta_test  = ml.load("T1", "test")

# 2. Fit + Predict
model = ml.methods.LightGBMRegressor(task="T1")
model.fit(X_train, y_train, seed=42)
y_pred = model.predict(X_test)

# 3. Score (cluster-bootstrap CIs by ticker for T1; adaptive n_boot)
metrics = ml.score("T1", y_test, y_pred)
print(metrics["mse"]["value"], metrics["mse"]["ci_lo"], metrics["mse"]["ci_hi"])

10 lines per task; swap the model class to compare methods.

Dataset structure

data/
├── daily/                     # primary granularity (4.84M panel rows)
│   ├── panel_train.parquet                       # T1, T4 train side
│   ├── panel_test.parquet                        # T1, T4 eval side
│   ├── scenarios.parquet                         # 1,130 macroeconomic events
│   ├── valuation_inputs.parquet                  # T2 features
│   ├── valuation_ground_truth.parquet            # T2 ground truth (market cap)
│   ├── private_valuation_inputs.parquet          # T5 features (price-stripped)
│   ├── private_valuation_ground_truth.parquet    # T5 ground truth
│   ├── generation_inputs.parquet                 # T3 fundamentals snapshot
│   ├── generation_ground_truth.parquet           # T3 long-form (ticker, FY, field, value)
│   ├── generator_eval_inputs.parquet             # T6 NL company descriptions
│   ├── generator_eval_ground_truth.parquet       # T6 long-form
│   └── scenario_forecast_ground_truth.parquet    # T4 ground truth
├── weekly/                    # Friday-close resampled (1.01M rows) — same file set as daily/
├── monthly/                   # Last-trading-day resampled (232k rows) — same file set as daily/
├── real_estate/
│   ├── re_train_properties.parquet   # T7 train (53,804 unique addresses)
│   ├── re_eval_inputs.parquet        # T7 eval (23,190 unique addresses)
│   └── re_eval_ground_truth.parquet  # T7 ground truth (rent + price)
├── xbrl/                      # 46.8M standardized XBRL facts, 92.6% ticker coverage
├── filings/                   # 295,860 SEC filings (10-K, 10-Q, 8-K, 20-F, 6-K, N-CSR, N-CSRS) — markdown + PDF
├── prices/                    # OHLCV + adjusted close (yfinance)
├── fundamentals/              # README placeholder only; raw CSVs reproducible via code/collect_fundamentals.py (see note below)
└── macro/                     # 46 FRED + 6 EIA series

manifest.json                  # SHA-256 over every parquet (provenance)

Note on data/fundamentals/: Raw per-ticker fundamentals CSVs (~12,000 files: {TICKER}_balance.csv, {TICKER}_income.csv, {TICKER}_cashflow.csv) are not shipped directly here because HF enforces a hard cap of 10,000 files per directory. They are reproducible end-to-end via the released pipeline:

python code/collect_fundamentals.py

The aggregated/processed versions used by the benchmark API (e.g. valuation_inputs.parquet, private_valuation_inputs.parquet, generation_inputs.parquet) are shipped directly under data/daily/, data/weekly/, data/monthly/, and data/real_estate/.

Data sources & access requirements

What's bundled in this HF release (no user credentials required):

Source	Bundled artifact	License
SEC EDGAR	filings/ (295k docs), xbrl/ (46.8M facts)	Public domain (US gov)
FRED	46 macroeconomic series	Public domain
EIA	7 commodity series	Public domain
yfinance	prices/ (OHLCV), fundamentals/ (quarterly) — derived features	Non-commercial (yfinance ToU)
RentCast	real_estate/ (address-level derived features only — rent + price targets, property attributes)	RentCast ToU — derived only
Macroeconomic events	scenarios.parquet (1,130 events × 49 types)	Curated by us, CC-BY-4.0

What's NOT bundled (gated — user credentials required for raw re-fetch via collect_*.py):

Source	Status	User-side requirement
Financial-news provider	Excluded — provider ToU prohibits redistribution. The release ships derived counts (filing_8k_count_30d, news_count_7d, has_press_release_7d) only.	User's own news-API key required for collect_news.py
RentCast raw listings	Excluded raw — proprietary. Derived features bundled.	User's own RentCast subscription required for collect_real_estate.py raw mode

Universe

The 4,416-ticker universe combines: full Russell 2000 (1,923 IWM holdings), full S&P SmallCap 600 (72 IJR-only additions), iShares Micro-Cap (225 IWC additions), and 2,196 small-cap NASDAQ/NYSE tickers outside all three indices. The split is 3,857 operating companies + 333 funds + 226 SPACs, with security_type recorded for applicability-aware stratification.

Splits

• Forecasting (T1, T4): chronological 70/30 split at 2024-09-03.
• Valuation + generation (T2, T3, T5, T6): 30% company-level holdout = 1,324 tickers (seed = 42), each contributing its latest valid snapshot.
• Real-estate (T7): 30% address-level holdout (random, seeded), with per-property time-axis features.

Cluster-bootstrap 95% CIs are computed per task: by ticker (T1/T2/T3/T5/T6), scenario_id (T4), or address (T7). Number of bootstrap resamples is adaptive in [1k, 10k] until (ci_hi - ci_lo) / |mean| < 0.05.

Methods (panel)

The release ships a 18-method baseline panel across 7 families: 4 naive, 2 classical, 3 deep sequence, 3 zero-shot TSFM, 2 LLM-adapted multi-task systems, 3 zero-shot frontier LLMs (gpt-oss-120b, gpt-5.1, gemini-3-flash + qwen35). Every method registers via @register(name=…, family=…, tasks=…) and exposes the sklearn-style (fit, predict, save, load) contract.

ml.list_methods()                 # all registered methods
ml.list_methods(task="T1")        # methods that support T1
ml.list_methods(family="naive")   # naive baselines

License

• Data: CC-BY-4.0 (derived features + curated panel)
• Code: MIT (code/macrolens/, code/dataloader/, code/methods/, code/eval.py, code/experiments/)
• Vendored libraries (under code/methods/_vendored/):
  • tslib/ — MIT (DLinear, iTransformer source)
  • moderntcn/ — Apache 2.0 (ModernTCN source)
• Reconstruction scripts (collect_*.py) provided for sources with redistribution restrictions: SEC filings (re-fetch from EDGAR), financial news (re-fetch from provider), real-estate (re-fetch from RentCast).

Citation

@inproceedings{macrolens2026,
  title = {{MacroLens}: A Multi-Task Benchmark for Contextual Financial Reasoning under Macroeconomic Scenarios},
  author = {<authors>},
  booktitle = {NeurIPS 2026 Evaluations & Datasets Track submission},
  year = {2026}
}

Reproducibility

When code/experiments/run_all.py is executed, each method run produces a RunRecord JSON under code/experiments/results/ recording: git_sha, lib_versions, hardware, artifact_sha256 (SHA-256 of every parquet read), timestamp, and deterministic_mode. Predictions are also persisted at code/experiments/predictions/<method>_<task>_seed<seed>.pkl so eval logic can be re-applied via code/experiments/re_evaluate.py without re-running the models. These output directories are not shipped on Hugging Face — reviewers reproduce them by running the released code.

Reconstruction (raw filings + news)

The release ships derived features and reconstruction scripts; raw artifacts subject to redistribution restrictions remain re-fetchable:

python collect_universe.py        # iShares ETF holdings + NASDAQ Trader directory
python collect_filings.py         # SEC EDGAR (10-K, 10-Q, 8-K, 20-F, 6-K, N-CSR, N-CSRS)
python collect_fundamentals.py    # XBRL company facts via SEC EDGAR
python collect_prices.py          # yfinance OHLCV + adjusted close
python collect_news.py            # provider-specific (~215k articles)
python collect_real_estate.py     # RentCast (100 metros, 139,855 properties)
python collect_macro.py           # FRED + EIA series
python preprocess.py
python assemble_benchmark.py
python generate_scenarios.py
python enrich_benchmark.py
python build_valuation_tasks.py
python validate_all.py

Authors / Contact

Anonymous (NeurIPS 2026 Evaluations & Datasets Track submission). Contact at <email> after author notification.