README.md

---
language: en
license: cc-by-4.0
task_categories:
- multiple-choice
- question-answering
size_categories:
- 1K<n<10K
pretty_name: FinDeepForecast-Historical-US
tags:
- finance
- economics
- macro
- forecasting
- federal-reserve
- time-series
- benchmark
dataset_info:
  features:
  - name: qid
    dtype: string
  - name: forecastType
    dtype: string
  - name: subtype
    dtype: string
  - name: indicator
    dtype: string
  - name: transform
    dtype: string
  - name: target_period
    dtype: string
  - name: info_cutoff
    dtype: string
  - name: forecast_end
    dtype: string
  - name: answer_release
    dtype: string
  - name: question
    dtype: string
  - name: options
    list: string
  - name: answer_letter
    dtype: string
  - name: answer_raw
    dtype: string
  - name: unit
    dtype: string
  - name: year
    dtype: int32
  splits:
  - name: '1999'
    num_bytes: 74833
    num_examples: 236
  - name: '2000'
    num_bytes: 104028
    num_examples: 328
  - name: '2001'
    num_bytes: 100621
    num_examples: 318
  - name: '2002'
    num_bytes: 107149
    num_examples: 337
  - name: '2003'
    num_bytes: 107172
    num_examples: 336
  - name: '2004'
    num_bytes: 101665
    num_examples: 322
  - name: '2005'
    num_bytes: 103728
    num_examples: 328
  - name: '2006'
    num_bytes: 101349
    num_examples: 321
  - name: '2007'
    num_bytes: 102494
    num_examples: 321
  - name: '2008'
    num_bytes: 113226
    num_examples: 352
  - name: '2009'
    num_bytes: 100911
    num_examples: 318
  - name: '2010'
    num_bytes: 106598
    num_examples: 335
  - name: '2011'
    num_bytes: 106190
    num_examples: 334
  - name: '2012'
    num_bytes: 105532
    num_examples: 333
  - name: '2013'
    num_bytes: 101793
    num_examples: 322
  - name: '2014'
    num_bytes: 104869
    num_examples: 331
  - name: '2015'
    num_bytes: 108792
    num_examples: 341
  - name: '2016'
    num_bytes: 113696
    num_examples: 356
  - name: '2017'
    num_bytes: 102019
    num_examples: 324
  - name: '2018'
    num_bytes: 111264
    num_examples: 348
  - name: '2019'
    num_bytes: 102219
    num_examples: 324
  - name: '2020'
    num_bytes: 107109
    num_examples: 334
  - name: '2021'
    num_bytes: 102620
    num_examples: 324
  - name: '2022'
    num_bytes: 109808
    num_examples: 343
  - name: '2023'
    num_bytes: 107952
    num_examples: 339
  - name: '2024'
    num_bytes: 108210
    num_examples: 339
  - name: '2025'
    num_bytes: 106355
    num_examples: 329
  - name: '2026'
    num_bytes: 49530
    num_examples: 155
  download_size: 994498
  dataset_size: 2871732
configs:
- config_name: default
  data_files:
  - split: '1999'
    path: data/1999-*
  - split: '2000'
    path: data/2000-*
  - split: '2001'
    path: data/2001-*
  - split: '2002'
    path: data/2002-*
  - split: '2003'
    path: data/2003-*
  - split: '2004'
    path: data/2004-*
  - split: '2005'
    path: data/2005-*
  - split: '2006'
    path: data/2006-*
  - split: '2007'
    path: data/2007-*
  - split: '2008'
    path: data/2008-*
  - split: '2009'
    path: data/2009-*
  - split: '2010'
    path: data/2010-*
  - split: '2011'
    path: data/2011-*
  - split: '2012'
    path: data/2012-*
  - split: '2013'
    path: data/2013-*
  - split: '2014'
    path: data/2014-*
  - split: '2015'
    path: data/2015-*
  - split: '2016'
    path: data/2016-*
  - split: '2017'
    path: data/2017-*
  - split: '2018'
    path: data/2018-*
  - split: '2019'
    path: data/2019-*
  - split: '2020'
    path: data/2020-*
  - split: '2021'
    path: data/2021-*
  - split: '2022'
    path: data/2022-*
  - split: '2023'
    path: data/2023-*
  - split: '2024'
    path: data/2024-*
  - split: '2025'
    path: data/2025-*
  - split: '2026'
    path: data/2026-*
---

# FinDeepForecast-Historical-US

A **historical** version of the [FinDeepForecast](https://openfinarena.com/fin-deep-forecast/) benchmark from OpenFinArena, covering **1999–2026** with ground-truth derived from real FRED time series.

Strictly follows the paper's two-track taxonomy:
- **Recurrent** — periodic numerical forecasts (CPI/GDP/Treasury/etc. value at a future date)
- **Non-Recurrent** — binary YES/NO forecasts on specific upcoming scheduled events (FOMC rate decisions, CPI/NFP release surprises, weekly market thresholds)

## Highlights

| Metric | Value |
|---|---|
| Total questions | **8,437** |
| Recurrent | 6,366 (75.5%) — multiple choice (4 options) |
| Non-Recurrent | 2,071 (24.5%) — binary YES/NO |
| Years covered | 1999–2026 (28 splits) |
| Indicators | 49 US macro/market series from FRED |
| Avg per year | ~300 questions |

## Quick Start

```python
from datasets import load_dataset

# Single year
ds = load_dataset("TheFinAI/pre_test", split="2008")

# Filter by forecast type
recurrent_2008 = ds.filter(lambda x: x["forecastType"] == "Recurrent")
non_recurrent_2008 = ds.filter(lambda x: x["forecastType"] == "Non-Recurrent")
```

## Recurrent (paper-aligned periodic forecast)

> *"Forecast the value of [US CPI YoY Inflation Rate] for June 2010.
> (Information available up to 2010-04-15.)"*
>
> A) 1.45%   B) 2.04%   C) 2.42%   D) 1.85%

- Format: 4-option MCQ with numeric values
- Generated for 49 indicators × 4 quarters/year × {level, yoy_pct, yoy_pp} transforms
- `info_cutoff` set ~60 days before target period

## Non-Recurrent (paper-aligned binary YES/NO)

Follows the original paper's format: "Will [specific event] happen by [date]?"

### 8 templates (all anchored to scheduled events)

| Template | Question pattern | Per year |
|---|---|---|
| T1 `fomc_cut` | Will FOMC cut rates ≥25bp at [date]? | ~8 |
| T2 `fomc_hike` | Will FOMC raise rates ≥25bp at [date]? | ~8 |
| T3 `fomc_hold` | Will FOMC keep rates unchanged at [date]? | ~8 |
| T4 `cpi_release_threshold` | Will CPI YoY for [month] exceed [threshold]%? | 12 |
| T5 `nfp_release_threshold` | Will NFP for [month] show change > [threshold]? | 12 |
| T6 `gdp_release_threshold` | Will GDP YoY for [quarter] exceed [threshold]%? | 4 |
| T7 `vix_weekly_spike` | Will VIX exceed [threshold] in week ending [date]? | 12 |
| T8 `nasdaq_weekly_gain` | Will NASDAQ gain more than [X]% in week ending [date]? | 12 |

### Non-Recurrent example (real)

> *"Will the FOMC cut the federal funds target rate by at least 25 basis points
> at its meeting on 2020-03-15?"*
>
> A) YES   B) NO
>
> ✓ Answer: A (YES) — Fed cut to zero on emergency Sunday meeting

### YES/NO balance (across all 2,071 NR questions)

| Subtype | n | YES% | NO% |
|---|---:|---:|---:|
| vix_weekly_spike | 328 | 51% | 49% |
| nasdaq_weekly_gain | 328 | 41% | 59% |
| nfp_release_threshold | 326 | 49% | 51% |
| cpi_release_threshold | 313 | 52% | 48% |
| fomc_cut | 224 | 17% | 83% |
| fomc_hike | 224 | 19% | 81% |
| fomc_hold | 224 | 64% | 36% |
| gdp_release_threshold | 104 | 47% | 53% |

(FOMC imbalance reflects reality: most meetings are "hold" decisions.)

## Schema

| Field | Type | Description |
|---|---|---|
| `qid` | string | Unique question ID |
| `forecastType` | string | `Recurrent` or `Non-Recurrent` |
| `subtype` | string | Fine-grained subtype |
| `indicator` | string | Primary FRED series ID |
| `transform` | string | `level` / `yoy_pct` / `yoy_pp` (Recurrent), `""` (NR) |
| `target_period` | string | Period asked about |
| `info_cutoff` | string | YYYY-MM-DD — latest info allowed |
| `forecast_end` | string | YYYY-MM-DD — last day of horizon |
| `answer_release` | string | NR only — when truth becomes verifiable |
| `question` | string | Question text |
| `options` | list[string] | Length 4 (Recurrent) or 2 (NR) |
| `answer_letter` | string | A/B/C/D (Recurrent) or A/B (NR) |
| `answer_raw` | string | Underlying answer value |
| `unit` | string | `%`, `index`, `binary`, etc. |
| `year` | int | Convenience field |

## Indicators (49 FRED series)

| Category | Examples |
|---|---|
| Inflation (8) | CPIAUCSL, CPILFESL, PCEPI, PCEPILFE, PPIACO, PPIFIS, DCOILWTICO, DCOILBRENTEU |
| Labor (6) | UNRATE, PAYEMS, CIVPART, EMRATIO, AHETPI, ICSA |
| Growth (4) | GDPC1, GDP, INDPRO, TCU |
| Rates (8) | FEDFUNDS, DGS3MO, DGS2, DGS5, DGS10, DGS30, T10Y2Y, MORTGAGE30US |
| Money (4) | M2SL, BOGMBASE, TOTBKCR, CCSA |
| Consumer (5) | UMCSENT, PCE, PSAVERT, RSAFS, DSPI |
| Housing (3) | HOUST, PERMIT, CSUSHPINSA |
| Manufacturing (3) | DGORDER, BOPGSTB, NEWORDER |
| Market (8) | SP500, NASDAQCOM, DJIA, VIXCLS, DTWEXBGS, DEXUSEU, DEXJPUS, DEXCHUS |

## Coverage Notes

- **1999** has fewer Recurrent questions (~180) because some indicators
  (`PPIFIS`, `SP500`, `DJIA`, `DTWEXBGS`) start later than 1999 on FRED.
- **2026** is a partial year (data through April/May 2026); contains only
  questions whose ground truth is verifiable.
- All NR questions tied to FOMC meetings, BLS releases (CPI/NFP), BEA releases (GDP),
  or weekly market thresholds — all from scheduled/public calendars.

## Differences from the Original FinDeepForecast

| Aspect | Original (live, 2025-10 → 2025-12) | This dataset (historical) |
|---|---|---|
| Coverage | 10 weeks | 28 years |
| Markets | 8 (US/UK/CN/HK/JP/SG/DE/FR) | US only |
| Recurrent total | 296 macro + 699 corporate | 6,366 (macro only) |
| Non-Recurrent total | 128 macro + 247 corporate | 2,071 (macro only) |
| Ground truth | Future outcome (live) | Historical realized values |
| Format | Numeric (Rec) + YES/NO (NR) | **Same** — 4-option MCQ + YES/NO |

This historical version sacrifices the live "no memorization" property of the
original benchmark in exchange for reproducible offline evaluation across 28 years.

## License

CC-BY-4.0. Underlying FRED data is in the public domain (FRED API terms).

## Citation

```bibtex
@article{findeepforecast,
  title={FinDeepForecast: A Live Benchmark for Financial Forecasting with LLMs},
  author={OpenFinArena},
  year={2026},
  url={https://openfinarena.com/fin-deep-forecast/}
}
```