Luminal Investment Review is a free, editorial publication. We read AI-investment studies and ask one question of each: could a second analyst reproduce this result from the data and code that were published? The pages below describe the three areas we examine on every review and the standard a study has to meet before we call its claim reproducible.
The three editorial areas
1. Data provenance
Every feature that feeds an AI-investment model has a history: a vendor, a public release, a collection window, a licence. We list that history for each input we can identify. A feature whose origin cannot be traced is recorded as unprovenanced rather than accepted on trust. Where a dataset has been revised or point-in-time contaminated — for example, fundamentals that were restated after the test window — we note it, because it changes what a backtest actually measured.
2. Reproducible pipelines
A claim earns the word reproducible only when the path from raw data to the published number can be run a second time. We check for fixed random seeds, pinned code versions, and a written record of each preprocessing step. Where a study shares its conclusions but keeps its pipeline private, we record that gap explicitly and label the result not independently reproducible rather than guessing from a clean-looking chart.
3. Model lineage
The model behind a claim is more than an architecture name. We record the family, the version, the checkpoint or weights used, and the training cutoff. When a backtest relies on weights that were later retrained or never specified, we treat it as a separate experiment from anything a reader could run today, and we say so in the note.
The standard a study must meet
Before we publish a verdict, a study has to clear four bars, in order:
- A specific claim. A stated return, accuracy, or prediction with a date and a source. Vague aspirations are not testable and we do not rate them.
- A retrievable dataset. Either public data we can fetch or a vendor source we can name. Data that exists only on a private server is rated as not reproducible.
- A runnable pipeline, or an honest gap. Where code is published we re-run it; where it is withheld, we record which pieces are absent rather than guessing at them.
- A clear verdict. Each note ends with one of three labels: reproducible, partial, or not reproducible, with the evidence behind it.
What we do not publish
We do not publish stock picks, portfolio recommendations, or performance forecasts. We do not rate the investment merit of a strategy — only the soundness of the evidence behind its claimed results. Nothing on this page or anywhere on this site is investment advice.
Suggest a study
If you have read an AI-investment claim you would like us to examine — a paper, a vendor deck, a backtest summary — send it through the inquiry form with the source and any linked dataset or code. We consider every suggestion for a future note.