Actual AI analysis · Qwen3-4B LLM
How AI analyses an earnings release.
This example comes from the current processing run. It shows what Qwen3-4B LLM records; it is not a prediction and was not selected based on subsequent share-price performance.
Salesforce · earnings release dated 28 February 2024 · SEC accession 0001108524-24-000002.
Model output
The financial interpretation at a glance.
The internal model score does not mean there is a 75% probability that the analysis or a price prediction is correct. Manual accuracy assessment is a separate step. Scale: −2 means strongly negative, −1 negative, 0 neutral or not mentioned, +1 positive and +2 strongly positive. Each field is assessed separately, so positive overall sentiment can coexist with a negative margin signal.
Traceable evidence
Non-neutral scores require supporting evidence.
The automated check looks for each evidence excerpt verbatim in the selected passages. In this example, every quoted excerpt could be located in the selected input (100%). This confirms where the evidence appears, not whether the interpretation is correct.
“Fourth Quarter Revenue of $9.29 Billion, up 11% Year-Over-Year”
“Share Repurchase Program Authorization Increased by $10 Billion”
From passage to field
Which statement supports each extracted feature?
“Fourth Quarter Revenue of $9.29 Billion, up 11% Year-Over-Year”
The quotation can be located in the source. Its interpretation has not yet been validated by a human reviewer.
“Share Repurchase Program Authorization Increased by $10 Billion”
An authorisation is not the same as a completed share repurchase.
Only the source release
The model receives selected financial passages and the publication timestamp, without subsequent price data.
Fixed JSON schema
Each analysis produces the same set of scores, quality fields, a summary and evidence excerpts.
One example is not a final conclusion
Only the complete holdout evaluation and independent replication can establish whether the features have predictive value.