Research instrument · v1

Can an LLM read the market?

xTrade puts language models on the record. Pick a company and a date, hand the model 56 days of news, filings and prices, and score its 7-day call against what actually happened.

Open the appSee how it works
OpenAIAnthropicGoogle GeminiGoogle VertexAlibaba QwenDeepSeekOpenAIAnthropicGoogle GeminiGoogle VertexAlibaba QwenDeepSeek
Evidence · window
  • NVDA supply constraints ease into Q4
  • Datacenter capex guidance raised
  • 10-Q filed · quarterly report
  • Analyst notes: sector rotation
Realized · 7d
Prediction · as-of
NVDA2025-08-14
Correct
Score · 7d trend
+1
up
-2-10+1+2
48 articles
3 filings
40 bars
+1 · up

The corpus, in numbers

0
companies tracked
0
news articles
0
SEC filings
0
daily price bars
0
runnable date · company pairs
0
sectors
0
LLM providers
0
day evidence window

56-day evidence window · 7-day prediction horizon · scores from -2 to +2.

The loop

Register, pick, prompt, settle.

One question, answered from real evidence: how well can a given LLM predict a stock's next seven days?

01

Register a model

Your provider, your model id, your key. OpenAI, Anthropic, Gemini, Vertex, Qwen, DeepSeek.

02

Pick a company and a date

We show only combinations with enough evidence to produce a valid run.

03

The model gets the evidence

56 days of news, filings and prices assembled into one prompt.

04

The market settles the score

Checked against the next 7 days of closes. Correct, incorrect, or still pending.

Coverage ruleEvery 7-day week in the window needs ≥ 7 linked articles and ≥ 2 with the company in the headline. Runs that don't clear the bar never get offered — no calls wasted on thin windows.
The instrument

A five-notch scale, from conviction to conviction.

Every prediction is a single integer between -2 and +2. Coarse on purpose — enough resolution to disagree, not so much that a model can hide in the decimals.

-1Down

The model leans bearish over the next 7 days.

-2-10+1+2

Click a notch · arrow keys to step

Traceability

Nothing is a black box.

Every score links back to the exact articles, filings, and price bars the model was shown. You can audit the input as easily as the output.

NVDAas of 2025-08-14 · 56-day window
48 articles3 filings40 bars
Articles
  • NVDA supply constraints ease into Q4
    Reuters · 2025-08-11
  • Hyperscalers raise datacenter capex guidance
    Bloomberg · 2025-08-08
  • Blackwell shipments track ahead of plan
    Financial Times · 2025-08-05
  • Analyst notes: sector rotation into semis
    WSJ · 2025-08-01
  • Sovereign AI orders extend into 2026
    The Information · 2025-07-28
Filings
  • 10-QQuarterly report · 2025-08-06
  • 8-KMaterial event · 2025-07-30
  • 10-KAnnual report · 2025-02-21
Price window
06-2007-1508-14
bars
40
Δ 56d
+12.4%
Comparison

Same company. Same date. Same evidence.

Different models. Now you can see who was actually right.

Leaderboard
NVDAas of 2025-08-14
Realized outcome · +1.4%
  • GPT-5OpenAI
    +1 Correct
  • Claude 4.5Anthropic
    +2 Correct
  • Gemini 2.5 ProGoogle
    0 Incorrect
  • Qwen 3 MaxAlibaba
    +1 Correct
  • DeepSeek V3.2DeepSeek
    -1 Incorrect

Illustrative product UI. Not presented as a claimed benchmark — run your own models in the app to produce real leaderboards.

Providers

Bring your own key.

Your credential, your quota, your call. Six providers today — add any model id they expose.

OpenAI
Anthropic
Google Gemini
Google Vertex
Alibaba Qwen
DeepSeek
Manual modeNo API? Copy the generated prompt into any chat LLM and paste the answer back — a model without an API can still be benchmarked.

Put a model on the record.

One company. One date. Fifty-six days of evidence. See what your model actually thinks — and whether the market agrees.