pre_test — Historical Financial Forecasting MCQ

A 12,838-question multiple-choice benchmark built from 49 US macro / market indicators (FRED, 1999–2026). Every correct answer is a real historical value independently re-derivable from raw FRED data — not model-generated. All questions are 4-option (random baseline = 25%).

The dataset is designed to separate genuine forecasting from memorized recall: by comparing a model's accuracy across years (especially 2026, which is past most models' training cutoff and acts as a clean baseline), and across "blind" vs "revealed" framings of the same shock event.

Composition

Two broad kinds of question:

• Predictable (condition = forecast): the next periodic value of an indicator, which a model can in principle extrapolate from the pre-cutoff trend.
• Unpredictable (condition = event / blind / revealed): outcomes that cannot be derived from the pre-cutoff trend, so above-25% accuracy on them signals memorized recall.

condition	count	predictable?	what it is
forecast	7,868	yes	periodic numeric forecast of an indicator (recurrent, quarterly cadence)
event	2,824	no	scheduled events with unpredictable outcomes — FOMC rate, NFP, CPI, GDP, weekly VIX peak, weekly NASDAQ return. Present every year as a stable floor.
blind	1,073	no	aftermath of a real black-swan shock, event name hidden — clean recall probe
revealed	1,073	no	the same shock question, event name given — situational reasoning

blind/revealed are paired 1:1 via pair_id (same options, same answer) and cluster in crisis years. The event floor guarantees every year has a ~25–30% unpredictable share (crisis years rise to ~50% once black swans stack on top). Answer letters are balanced: A/B/C/D ≈ 25% each. The recurrent base is ~280–312/year (1999 and 2026 are lighter — fewer indicators existed in 1999; 2026 is partial).

Fields

Input to the model:

• question — the full prompt, options already appended.
• options — the four "A) …" strings (also for programmatic use).

The answer:

• answer_letter — "A"/"B"/"C"/"D".
• answer_raw — the numeric value of the correct option (e.g. "+0.32pp").

Metadata / analysis labels (NOT shown to the model):

• condition — forecast / event / blind / revealed (the experiment types above).
• subtype — recurrent / shock_aftermath / event subtypes (fomc_rate, nfp_change, cpi_yoy, gdp_yoy, vix_weekly, nasdaq_weekly).
• forecastType — Recurrent / Non-Recurrent.
• indicator — FRED series id (e.g. CPIAUCSL).
• transform — level / yoy_pct / yoy_pp.
• target_period, info_cutoff — what is being predicted, and the information cutoff (strictly before the target → it is forecasting, not lookup).
• unit, year.
• pair_id, event_label, event_date — present on shock rows; link the blind/revealed twins.

Recommended analysis

1. Accuracy vs. year — group by year; a cliff at 2026 (→ ~25% chance level) is the clean signal that pre-2026 accuracy is partly memorization.
2. Blind vs. revealed — join on pair_id; a gap suggests the model relies on the event name (memory) rather than the pre-cutoff numbers (reasoning). (Auxiliary signal — note the date itself can leak event identity.)

Loading

from datasets import load_dataset

# all rows
ds = load_dataset("lfqian/pre_test", split="train")

# one year at a time (config "by_year", splits y1999 … y2026)
ds_2008 = load_dataset("lfqian/pre_test", "by_year", split="y2008")

# every year as a DatasetDict
by_year = load_dataset("lfqian/pre_test", "by_year")

Files

• mcq_all.parquet — all rows (default config, train split).
• mcq_all.jsonl / mcq_all.json — same data, JSONL and single-array JSON.
• mcq_<year>.parquet — per-year files (config by_year, splits y1999…y2026).
• mcq_<year>.json — per-year files in JSON.

Provenance

Indicators and values from the Federal Reserve Economic Data (FRED). Questions, distractors, and answer-letter balancing generated by the OpenFinArena competitions_historical pipeline. 100/100 stratified audit re-derived every sampled answer from raw FRED.