Blame for README.md

---
license: cc-by-4.0
language:
- en
pretty_name: FinDendro
size_categories:
- 10K<n<100K
task_categories:
- question-answering
- multiple-choice
tags:
- finance
- economics
- knowledge-cutoff
- temporal-knowledge
- llm-evaluation
- look-ahead-bias
- backtesting
configs:
- config_name: questions
  data_files: data/questions.jsonl
  default: true
- config_name: sample_facts
  data_files: data/sample_facts.jsonl
- config_name: fact_pool_macro_policy
  data_files: data/fact_pool_macro_policy.jsonl
- config_name: results_frontier
  data_files: results/frontier/*.jsonl
- config_name: results_open
  data_files: results/open/*.jsonl
- config_name: results_point_in_time_letter
  data_files: results/point_in_time/letter/*.jsonl
- config_name: results_point_in_time_likelihood
  data_files: results/point_in_time/likelihood/*.jsonl
---

# FinDendro: Temporal Self-Knowledge of LLMs on Dated Financial Facts

FinDendro measures two dates for a language model: where its knowledge of dated financial facts actually ends, and where the model *says* it ends. Every answer is tied to the model's own claim of knowing it, so the benchmark can test the safeguards used in LLM backtests (vendor-stated cutoffs, self-reports, self-report filters, and "assume today is ..." prompts).

## Facts

Dated, effectively unrevised U.S. facts from January 1999 to September 2026, in three families:

| Family | Series | Source | Sample | Pool released here |
|---|---|---|---|---|
| Macroeconomic | 12 monthly series (headline/core CPI YoY, unemployment, effective fed funds, 3M/2Y/10Y Treasury, 30Y mortgage, JPY/USD, CNY/USD, USD/EUR, WTI) | FRED | 1,353 | 3,993 |
| Stock prices | Month-end close of 50 large U.S. companies, as quoted at the time (later splits reversed) | Public historical quotes | 540 | not released (see below) |
| Fed decisions | Monthly target-rate outcome (raised / lowered / unchanged, incl. intermeeting moves) | FRED DFEDTAR / DFEDTARU | 333 | 333 |

The evaluation sample has **2,226 facts** (sparse before 2021, dense from 2021 on) and **11,130 prompts** (five per fact). 1,340 facts count toward knowledge after the controls below.

## Questions (`data/questions.jsonl`)

Each fact yields five prompts that share one set of candidates. Numerical facts offer four values spaced by roughly one typical monthly change, with the true value's rank drawn uniformly; Fed decisions offer the three outcomes.

| `id` suffix | Task | Description |
|---|---|---|
| `/T2` | Task 2 (value) | Choose the value. Reference for what the model knows. |
| `/T1-recall-none` | Task 1, Recall | Every candidate offered twice: "I know this value" or "I do not know this value; my best guess is". |
| `/T1-recall-declared` | Task 1, Recall + declared date | Same, prefixed with "Assume that today is [end of the previous month]." |
| `/T1-predict-none` | Task 1, Predict | "I already know the outcome" or "I do not know the outcome; my prediction is". |
| `/T1-predict-declared` | Task 1, Predict + declared date | Same, with the declared date. |

Key fields: `prompt`, `gold` (option letter), `option_rank` (letter → candidate rank; for Task 1 see `option_meta` in the results, which gives `stance` and `rank`), `true_rank`, `persistence_correct` (True if repeating last month's value would land on the correct option), `family`, `period`, `indicator`.

**Controls used in the paper.** Numerical knowledge is measured only on facts with `persistence_correct == False`; Fed decisions use balanced accuracy over the three outcomes ("unchanged" is correct in 78% of months). Chance accuracy is about 0.27.

## Raw model results (`results/`)

One JSON line per prompt, with the model's raw `response`.

| Folder | Models | Scoring |
|---|---|---|
| `frontier/` | GPT-5, GPT-4o, DeepSeek-V3.2, Llama-3.3-70B-Instruct (API, temperature 0 where supported) | single option letter |
| `open/` | Qwen3-8B, Qwen3.5-9B (greedy, thinking disabled) | single option letter |
| `point_in_time/letter/` | TiMaGPT2, ChronoGPT-Instruct, DatedGPT vintages | single option letter (these small models mostly pick option A) |
| `point_in_time/likelihood/` | the same families, Task 2 only | candidate with the highest mean log-probability (`value_scores`) |

`results/summary/` holds the aggregate metrics: `headline_metrics.json` (knows, says, accuracy on claimed/disclaimed facts, within-period gap, declared-date effects), `quarterly_timeline.json` (quarterly accuracy and claim rate, and the estimated knowledge-end and claims-end quarters), and `extra_analyses.json` (by fact family, Recall vs Predict wording, self-report filter, threshold robustness).

## Headline results

| Model | Knows | Says it knows | Acc. on disclaimed facts | Knowledge ends | Claims end | Vendor cutoff |
|---|---|---|---|---|---|---|
| GPT-5 | 0.76 | 0.38 | 0.61 | 2024Q3 | 2024Q3 | 2024-09 |
| DeepSeek-V3.2 | 0.54 | 0.28 | 0.48 | 2024Q3 | 2024Q4 | n/a |
| GPT-4o | 0.50 | 0.20 | 0.34 | 2023Q1 | 2023Q2 | 2023-10 |
| Llama-3.3-70B | 0.43 | 0.48 | 0.29 | 2021Q2 | 2023Q1 | 2023-12 |

## Notes and licensing

- Questions, annotations and model outputs are released under CC BY 4.0. Model outputs remain subject to the respective providers' terms.
- Macroeconomic and Fed series come from FRED; please also respect the terms of the original data providers listed on FRED.
- Stock prices are released only as question candidates and answers. The full monthly price series used to build the pool is not redistributed.
- Model outputs were collected in October 2026. API-served models change over time, so re-running may give different answers.