THE PROJECT IN ORDINARY WORDS
A system that looked twice as good as it was.
This project ranks newly filed SEC disclosures by how likely each is to matter. The interesting part is not the ranking. It is that an early version scored roughly twice as well as the honest one, the reason was a single assumption about a timestamp, and the difference was measured and published rather than quietly fixed.
The problem
Too much is filed to read
Public companies file continuously, and most of what they file changes nothing. A person who wants to keep up has to decide, every morning, which of the day's filings are worth opening — and that decision is usually made by headline, by company size, or not at all.
So the question this project asks is narrow and checkable: given everything filed today, which five should be read first? A ranking can be scored against what actually happened afterwards, which is why it was chosen over anything that would require predicting a price.
The evaluation covers 9,721 filings over 957 sessions, a median of 9 filings per session.
Where the data comes from
Three sources, each with a time of observation
SEC filings
8-K disclosures from EDGAR, with the acceptance timestamp kept — the moment the filing became public, not the date printed on it.
Daily prices
Adjusted daily open, high, low and close, used to measure what happened after a filing and to build the historical pictures on company pages.
Company Facts
Filed financial figures with their accounting concept, unit and reporting period, used for the filed-results section rather than for the ranking model.
Nothing is bought, scraped from a terminal, or hand-entered. Every input is a public filing or a vendor-adjusted price series, and the evaluation records the row count and content hash of each.
What it does and does not do
The boundary, stated
What it does
- Ranks the day's filings by the estimated chance of an unusually large benchmark-adjusted move afterwards.
- Measures that ranking against what happened, on sessions the model never trained on.
- Publishes the measurement, the failures and the evidence files behind both.
What it does not do
- It does not give investment advice, price targets or buy and sell calls.
- It estimates the size of a reaction, never its direction. A large move can go either way.
- A historical ranking result is not a trading return. Nothing here accounts for costs, slippage, capacity or the price of being wrong.
What it scored
What it scored, honestly measured
What that number is. It is the average share of each day's top five that turned out to be material. It is not the chance of finding at least one material filing in the top five, and it is not a success rate per day. On a typical session with five picks, it means roughly one of them mattered.
The 95% interval on average precision runs 36.5% to 43.0%, from a cluster bootstrap over sessions with 2,000 draws. A point estimate without that interval would suggest more precision than the sample supports.
The timestamp
One assumption about a timestamp
A filing is accepted at a particular minute, and a reader can only act after it. The first pipeline entered positions on the filing *date* instead — which, for a filing accepted after the close, means entering before it existed. Nothing else about the model changed.
Measured on the same sessions, the shortcut scored 31.2% where the honest pipeline scores 19.8%. Each row below removes one more thing the model should not have known.
| Stage | Daily top-five share | Average precision | Entries that were impossible |
|---|---|---|---|
| Naive pipeline | 31.2% | 59.7% | 96% |
| + purged, embargoed CV | 31.2% | 59.0% | 96% |
| + shifted trailing features | 27.4% | 43.9% | 96% |
| + point-in-time universe | 27.4% | 43.9% | 96% |
| + point-in-time entry | 19.8% | 39.7% | 0% |
What did not work
Three things that were tried and not shipped
Written as a list of negatives on purpose. All three were measured with the same machinery as the positive results and reported the same way.
Language features from the filing text
Sentiment and change-detection features over the filing body did not improve the ranking beyond what arrival time and item code already carried.
Issuer-relative scoring as a ranking signal
Comparing a filing with its own issuer's history is useful on a company page, and it did not improve the cross-sectional daily ranking.
Volatility forecasting
A forecaster only appears on a company page when it passes its out-of-sample coverage gate. When none does, the card is absent rather than approximate.
What is running
What is running
The research engine refreshes filings, prices and filed results on a schedule and regenerates every company page from them. The site you are reading is the output, and each number on it links to the filing or price series behind it.