Examining how AI investment claims hold up against data provenance and reproducible research — one method, one dataset, one published note at a time.
Read the research focus
The provenance problem
A model that picks stocks is only as trustworthy as the prices, fundamentals, and alternative data it was trained on. Yet the trail from a published return figure to the underlying dataset is rarely published. We start every review by asking where the numbers came from, how they were cleaned, and whether anyone else could rebuild the same experiment.
That question — could a second analyst reproduce this? — is the spine of every article we publish here.
What we examine
Each note follows the same inspection path so readers can compare studies on equal terms.
01 · Data provenance
We map every claimed input — price feeds, fundamentals, news sentiment, satellite or alternative data — to its original source, licence, and collection window. A feature with no retrievable origin is flagged, not quietly accepted.
02 · Reproducible pipelines
We look for fixed seeds, versioned code, documented preprocessing, and a runnable notebook or script. When a study only publishes conclusions and not the pipeline, we say so plainly and rate the claim as not independently reproducible.
03 · Model lineage
From the architecture family to the trained weights, we record what was actually evaluated. A backtest run against an unspecified or later-retrained model is treated as a different experiment from the one a reader would face today.
Our editorial method
A short, repeatable process keeps every article honest and comparable.
We quote the return, accuracy, or prediction figure exactly as published, with its date and source, before any analysis begins.
We list every dataset behind the claim, its vendor or public origin, and the time window used for training and evaluation.
Where code or data is available, we re-run the pipeline from scratch. When it is withheld, we list precisely which pieces are missing instead of filling the gaps ourselves.
We release a short editorial note with the verdict — reproducible, partial, or not reproducible — and the evidence behind it.
A reproduction notebook is the artefact we look for first — and the one we publish when we can.
Who this is for
Questions we hear
No. The articles are editorial and informational only. Nothing published here recommends any particular security, and no personalised advice is offered to any reader.
No. There is nothing to buy here — no subscriptions, managed accounts, bookings, deposits, or trade execution. The editorial content is free; the only way to reach us is the inquiry form.
We pick AI-investment studies whose claims are specific enough to test — a quoted return, an accuracy figure, or a concrete prediction — and where some trace of data or code is public enough to follow.
Yes. Send the claim, its source, and any linked dataset or code through the inquiry form and we will consider it for a future note.
Have an AI-investment claim you want examined for provenance and reproducibility?
Send us an inquiry