AI-powered market intelligence
AI-driven financial research.
Feronix is an AI-driven financial research platform. We use language models, machine learning and statistics to investigate whether company announcements and market data help predict future share-price reactions. Independently tested, with uncertainty made explicit.
- Published studies
- 5
- Largest dataset
- 163,962 one-minute observations
- Research principle
- Out-of-sample always a separate final test
What does the AI do?
From company announcement to testable insight.
AI reads the text. A separate model investigates its relationship with share-price reactions. New data determines whether that relationship holds up.
- 01
Company announcement
An official earnings release with a public timestamp.
- 02
AI extraction
Qwen3-4B LLM reads financial passages and records supporting excerpts.
- 03
Financial features
Information on revenue, margins, outlook and risk is structured into features.
- 04
Prediction model
Features are linked to subsequent share-price reactions and compared with a simple baseline.
- 05
Independent final test
A separate test set assesses predictive value, accounting for uncertainty and costs.
Ongoing AI research · earnings releases
AI is reading 6,493 earnings releases. We test what it adds.
For each release, Qwen3-4B LLM extracts financial developments, risks, guidance and supporting excerpts. Only after independent replication will we assess whether these richer language features better explain share-price reactions.
View the study design and live progress- Processed
- Loading
- Model
- Qwen3-4B LLM
- Current phase
- Loading
- Latest stock
- —
- Time remaining
- Calculating
- Last updated
- —
Research journey
The research journey in five steps.
-
01
Initial study
607 releases
Interpretable text features: 55.93% directional accuracy. An indication, not yet evidence.
-
02
Independent replication
1,257 releases
49.88% against a baseline of 52.90%.
-
03
Conclusion
Not confirmed
The earlier advantage did not reappear on new data.
-
04
Currently active
Qwen3-4B LLM
6,493 releases are being reanalysed in greater depth.
-
05
Next
Locked-down final test
Results are published only after the prespecified evaluation.
Featured research
Read the full studies.
Independent replication
Did the earnings-release text signal hold up?
The model was locked down and tested on 296 new companies. The earlier advantage did not reappear.
- Releases
- 1,257
- Directional accuracy
- 49.88%
- Baseline
- 52.90%
One-minute data
Can oil predict Shell’s share price?
Four time horizons and 163,962 matched one-minute observations.
Best net result −11.74% 03SEC Form 4
Can clusters of insider purchases predict returns?
A reserved test period with 289 decision points.
Directional accuracy 49.13%Our standard
An attractive pattern is not yet evidence.
Many models look strong on the data used to develop them. That is why we assess results only after a chronological holdout, realistic costs and—where possible—a second dataset.
See how we testInteractive demonstration · keyword rules
From financial announcement to an interpretable example score.
Paste a fictional announcement or choose an example. This demo shows how simple keyword rules turn text into indicators and an illustrative direction, without a misleading confidence percentage.
Please note: this is a transparent browser simulation of the analysis steps—not live Qwen3-4B LLM output, a research finding or trading advice.
Summary of the main tone and uncertainties.
Feronix
Transparent about methods and limitations.
We combine financial data, local language models and reproducible analysis. Not trading signals, but verifiable answers to clearly defined research questions.
Defined in advance
The question, horizon and evaluation are fixed before the final test.
Tested with point-in-time data
A model sees only information available at the time.
Fully reported
Null results and limitations are reported prominently too.
Explore further
A clear path through the research.
Start with the conclusions, explore the methodology or see how the analysis is built.
What do we know so far?
Explore published studies, ongoing processing and results in context.
Results dashboard → Live research log → Documents & downloads →How do we test it?
From data sources and chronological ordering to model validation and interpreting results.
Learn to read results → Model validation → Technical architecture →Explore the steps yourself.
Inspect a source-backed Qwen extraction or compare existing studies in the Model Lab.
Analysis example → Model Lab → Frequently asked questions →Behind Feronix
Independent research, personally driven.
Floris Kieviet develops Feronix with AI assistants and a critical research mindset. Strong ideas must stand up to new data.