HeliumTrades/helium-market-resolution-benchmark

What is this?

Market Resolution contains 299 ranked option-contract questions. Most test calculations or comparisons from frozen quotes. Sixty test an implied-volatility prior with the premium hidden, and 11 test probability-of-finishing-in-the-money forecasts against a later outcome.

It does not measure trading profitability. It measures bounded option-chain reasoning: implied volatility (IV), delta, time value, parity, term structure, relative IV, chain surfaces, and P(ITM) forecasts.

Right now grok-4.20-reasoning leads with 58 mean points out of 100. That is a custom point score, not percent correct.

The short version: the best observed run still misses 42 of every 100 available points, and IV is the weakest major skill. A model may be useful for explanation, but these results do not support treating one as the calculator of record when money is at risk.

Why this matters to a person

A standard U.S. equity option contract usually controls 100 shares. Reading delta as 0.25 instead of 0.55 changes a simple delta hedge from roughly 25 shares to 55. Misreading time value changes how much of the premium appears to be payment for possibility rather than intrinsic value.

Those are concrete risk and cost errors. This benchmark asks the prerequisite question before advice or automation: can the model recover the relevant quantities from a fixed set of facts?

Use it to: screen a model for bounded option-chain calculations, or see whether a new release improved at chain math. It is not evidence that a model can trade profitably.
Items · Full results · Methodology

Who is ahead?

How to read this chart

• Each bar is mean points across 299 ranked questions. Five diagnostic checks sit outside the leaderboard.
• This is item-weighted, not task-weighted: four 60-item families contribute 240/299 questions.
• Multiple choice scores 0 or 1. Numeric scores decline linearly with error under published tolerances; P(ITM) uses 1 minus Brier score.
• Thin lines are 95% item-bootstrap intervals. They omit repeated-run variability and dependence among questions built from the same contract; they are not significance tests.

Human takeaway: every complete run leaves at least 42 of 100 available points unearned, and similar totals can hide different weaknesses. Check the task breakdown before trusting an overall rank.

Where do models earn points?

The story: IV families score lowest under their tolerance scales. P(ITM) scores highest, but it uses a different 1-minus-Brier scale with a 75-point uniform-random reference.

This chart shows all nine ranked task families. Task scales differ, so the bars are not directly comparable as percent correct. n is the number of ranked questions in that skill; each complete model run counts once.

If your use case depends on IV, the overall leaderboard is a poor substitute for the IV rows.

Example question

As of 2026-06-10, NVDA spot is $198.50. Contract: CALL strike 200, expiration 2026-06-18 (8 DTE).
Without using the option premium, estimate implied volatility (%). Reply with a single number.

This example tests an IV prior. Without an option premium or another volatility input, IV cannot be uniquely derived from spot, strike, and time.

Data caveat: only the NVDA/SPY high-volatility caches are source snapshots. Other ticker/date surfaces are documented transforms. Results are not evidence across six independent live markets; see Assumptions.

Does showing the market price help?

Same contracts, one evaluation: once with the premium hidden, once with bid/ask shown. Positive bars mean the price-shown prompt scored higher; negative bars mean it scored lower. No repeated-run or paired uncertainty is shown, so this is not a causal estimate.

Why care: in this run, giving the model more market information did not reliably improve its answer. More context can help, do nothing, or coincide with a worse score.

All scores, skill heatmap, scoring rules, load

All models

Model	Mean points	IV (no price)	IV (price shown)	Delta	Time value	Status
grok-4.20-reasoning	58	37	53	58	60	complete
grok-4.3	55	38	43	59	60	complete
claude-fable-5	54	52	36	50	60	complete
gpt-5.4	52	32	22	62	70	complete
qwen3-235b	51	43	38	59	55	complete
claude-opus-4.7	49	38	31	63	58	complete
claude-opus-4.6	47	29	25	65	67	complete
claude-4.5-sonnet	46	27	32	57	66	complete
gpt-5.6-sol	45	21	23	42	60	complete
gpt-5	44	34	25	53	53	complete
claude-sonnet-4.6	44	16	19	63	58	complete
llama-3.3-70b	37	36	20	61	38	complete
deepseek-v4	3	0	0	0	0	complete
gemini-3.1-pro	n/a	30	59	54	60	incomplete

Gemini 3.1 Pro finished 268/304 questions after repeated API timeouts. Its partial score is shown as incomplete and is not ranked against complete runs.

Green = stronger on that skill. Red = weaker. This is where similar overall scores come apart.

Scoring

• Multiple choice: right / wrong
• IV score falls linearly to zero at 14 to 18 percentage-point error; delta at 0.18 to 0.22 error
• Time-value tolerance is the larger of $0.15 or 15% of truth; P(ITM) uses 1 minus Brier
• Leaderboard = mean across the main item set; the skill split is the useful part

Load

from datasets import load_dataset
ds = load_dataset("HeliumTrades/helium-market-resolution-benchmark", split="test")

304 public items · 13 complete model runs + 1 incomplete · MIT · METHODOLOGY · ASSUMPTIONS
Sister benchmark: Model Worldview

@dataset{helium_market_resolution_2026,
  title={Helium Market Resolution Benchmark},
  author={Helium Trades}, year={2026}}