alinakhay/earnings-input-reliance-benchmark

Earnings Input Reliance Benchmark

Author: Alina Khay. Version: 1.0.0. Companion manuscript: Momentum, Commentary Tone, and Earnings Surprise in LLM Financial Forecasts, October 2026. No arXiv identifier is assigned here.

A controlled, synthetic benchmark measuring how financial information changes the answer preferences of six pinned Q8 instruction/chat checkpoint conversions. It independently reverses 20-day momentum, explicitly non-informational opinion tone, and whether fixed actual EPS beats manipulated consensus.

Configurations

Configuration	Rows	Unit
prompts	2,016	One unique expanded prompt/answer encoding
predictions	12,096	One expanded prompt evaluated by one conversion
primary_pairs	2,304	One model/scenario, with A/B mean and individual encodings
pilot_bridge	96	Selected old-pilot prompts rescored in the Q8 setup
blocks	24	One numerical construction block

All configurations use a test split as a repository convention. There is no training split or independently sampled market test set. Do not count repeated answer mappings or counterfactual evaluations as independent financial events.

Load

from datasets import load_dataset

prompts = load_dataset("alinakhay/earnings-input-reliance-benchmark", "prompts", split="test")
scores = load_dataset("alinakhay/earnings-input-reliance-benchmark", "primary_pairs", split="test")

Construction and provenance

Seed 20261008 generates 24 blocks crossing profit/loss, annual EPS level increase/decrease, 2%/15% gaps relative to absolute actual EPS, and three replicates. There are six positive/negative commentary pairs, three introductions, and two section orders. Each of 384 primary semantic scenarios has both A/B assignments and U/D. Additional cases test beat/miss recognition, commentary removal, natural-word answers, and an irrelevant office sentence.

The original protocol contains 1,824 prompts; an exploratory coverage amendment adds 192 profitable word/office cases. The amendment leaves primary, diagnostic, and neutral prompts unchanged. protocol.json, control_coverage_amendment.json, and original byte-preserved prompt files document that chronology. The expansion was internally frozen after inspection of a prior pilot, not publicly preregistered.

ChatGPT/Codex assisted benchmark construction, implementation, and documentation. Cases and commentary are synthetic authored variants, not independent human sentiment annotations or real company announcements. Recorded scores are actual local checkpoint evaluations, not invented model responses.

Models and score meaning

The six conversions are Qwen2.5-0.5B/1.5B/7B-Instruct, SmolLM2-1.7B-Instruct, TinyLlama-1.1B-Chat-v1.0, and Phi-3.5-mini-instruct. checkpoint_manifests.json identifies exact conversion/tokenizer revisions and file hashes. No weights are distributed and no model was fine-tuned for this experiment.

score is exp(positive-candidate log likelihood) divided by the sum for the two requested candidate prefixes. In numerical diagnostics the positive candidate means "beat"; in forecasts it means a positive reaction. score_ab_mean averages the two letter assignments. It is not a calibrated stock-return probability or a probability of a complete generated answer. Candidate-prefix mass is retained to expose conditioning on unlikely requested formats.

known_answer is present only for the diagnostic EPS comparison. Forecast gold_return_label is always null: no observed market returns are provided. schema.json lists columns and row counts. Raw predictions retain input/context, candidate token IDs, genuine log probabilities, prompt/token hashes and top-token information. summary.json provides the manuscript's derived statistics.

Intended use and limits

Use for matched input-sensitivity evaluation and answer-format robustness checks. Group counterfactuals and label reversals by numerical block in any held-out analysis. The six authored lexical pairs limit linguistic generalization. Short synthetic disclosures, fixed semantic-option ordering, answer-prefix conditioning, and conversion/precision sensitivity constrain interpretation. The selected pilot bridge is a setup comparison, not a random quantization-error sample. These data cannot establish trading profitability, realized forecast accuracy, internal reasoning, or precision-invariant properties of base models.

License and citation

Original benchmark data and result tables: CC BY 4.0; see LICENSE. Upstream checkpoint and tokenizer licenses remain separate.

Suggested citation: Alina Khay (2026), Momentum, Commentary Tone, and Earnings Surprise in LLM Financial Forecasts: A Controlled Study of Six Q8 Checkpoint Conversions. Cite the public paper identifier when assigned and the benchmark version used. No DOI or public paper identifier is invented in this release.

The accompanying manuscript remains copyright Alina Khay; the software/data licenses do not select a paper-submission license on her behalf.

Published resources

• Benchmark dataset
• Recorded-results explorer