Paper research. No broker. A rule that looks good here is still not an order.
The method is multiple regression: one equation that estimates how several measurements, together, lined up with the price change over the next five sessions. The measurements were written down first. The last stretch of history, 2026-04-07 through 2026-09-25, was not used to fit the equation. That later stretch is the only money test.
A neural net was set aside. On a few years of daily prices it has enough spare flexibility to memorize accidents, and it does not hand back a sentence you can audit. The useful test is the same either way: does the rule still work on a later period, after the cost of getting in and out. The simpler equation is the one that can fail in public.
There are 4 equations. One uses only price measurements. One uses only whether a contract or an earnings filing arrived. One puts those together. A fourth scores the wording of the Form 8-K itself. Fit period: 1800 non-overlapping rows. Later period: 575 rows. The share of later price moves lined up with the equation, in the fit period, is R². The earlier price equation’s R² was about 0.6%, which means almost all of the move was left unexplained. Paper budget for each traded rule: $250,000. Holding the same stocks over the later period, with the same kind of spread, finished at $292,909.
Recent return, a longer recent return, jumpiness, and unusual dollar volume.
In the fit period the equation lined up with 0.6% of the variation across 1800 rows. That percentage is a description of the past, not a forecast.
| Measurement | Coefficient | Rough size |
|---|---|---|
| baseline later return, when every measurement is at its usual level | +0.24 pp | +1.4 |
| return over the prior 5 sessions | +0.08 pp | +0.4 |
| return over the prior 20 sessions | -0.45 pp | -2.2 |
| jumpiness over the prior 20 sessions | +0.27 pp | +1.5 |
| today's dollar volume versus its recent usual level | -0.31 pp | -1.8 |
Not kept. Later period, $250,000 paper: finished at $239,478 (-10,522). 92 entries. -0.12% average trade after the spread. Holding the same stocks finished at $292,909. This book follows the measurements, not the ordinary drift: it buys up to four names when those measurements point to a return above the fit-period baseline.
A contract or closed deal in the last three sessions, and an earnings filing in that same window.
In the fit period the equation lined up with 0.0% of the variation across 1800 rows. That percentage is a description of the past, not a forecast.
| Measurement | Coefficient | Rough size |
|---|---|---|
| baseline later return, when every measurement is at its usual level | +0.24 pp | +1.4 |
| a contract or closed deal filed in the last 3 sessions | -0.08 pp | -0.5 |
| an earnings filing in the last 3 sessions | +0.01 pp | +0.0 |
No single measurement cleared the rough-size bar of 2, so none became its own rule.
Not kept. Nothing in the fit was large enough to call the later finish a rule. Later period, $250,000 paper: finished at $295,750 (+45,750). 92 entries. +0.80% average trade after the spread. Holding the same stocks finished at $292,909. This book follows the measurements, not the ordinary drift: it buys up to four names when those measurements point to a return above the fit-period baseline.
The multiple regression: all six measurements in one equation.
In the fit period the equation lined up with 0.6% of the variation across 1800 rows. That percentage is a description of the past, not a forecast.
| Measurement | Coefficient | Rough size |
|---|---|---|
| baseline later return, when every measurement is at its usual level | +0.24 pp | +1.4 |
| return over the prior 5 sessions | +0.09 pp | +0.5 |
| return over the prior 20 sessions | -0.45 pp | -2.2 |
| jumpiness over the prior 20 sessions | +0.26 pp | +1.5 |
| today's dollar volume versus its recent usual level | -0.34 pp | -1.8 |
| a contract or closed deal filed in the last 3 sessions | -0.10 pp | -0.5 |
| an earnings filing in the last 3 sessions | +0.08 pp | +0.4 |
Not kept. Later period, $250,000 paper: finished at $275,462 (+25,462). 92 entries. +0.50% average trade after the spread. Holding the same stocks finished at $292,909. This book follows the measurements, not the ordinary drift: it buys up to four names when those measurements point to a return above the fit-period baseline.
Opinion, agreement versus dispute, certainty, hard numbers, and promotional adjectives, scored from the company’s own 8-K with fixed word lists. This equation was added after the first three had already been scored on the same later period, so a win here is weaker evidence. Outside magazines and video transcripts are not in this test: they are not stored with a reliable clock.
In the fit period the equation lined up with 0.4% of the variation across 1800 rows. That percentage is a description of the past, not a forecast.
| Measurement | Coefficient | Rough size |
|---|---|---|
| baseline later return, when every measurement is at its usual level | +0.24 pp | +1.4 |
| a filing with enough text to score in the last 3 sessions | -1.65 pp | -1.6 |
| how much the filing judges, rather than only reports | +0.29 pp | +1.5 |
| consensus language minus dispute language in the filing | +0.05 pp | +0.2 |
| definite language minus hedging in the filing | -0.02 pp | -0.1 |
| how many hard numbers the filing uses | +0.10 pp | +0.2 |
| how promotional the filing’s adjectives are | +1.60 pp | +2.2 |
Kept as a paper research result. Later period, $250,000 paper: finished at $489,759 (+239,759). 29 entries. +3.95% average trade after the spread. Holding the same stocks finished at $292,909. This book follows the measurements, not the ordinary drift: it buys up to four names when those measurements point to a return above the fit-period baseline.
Each company’s words are the six most repeated content words in the annual report filed before 2025-09-25. Attention is how often people read the matching English Wikipedia pages in the prior 28 days, compared with the 84 days before that. The news archive GDELT refused the connection from this network, so this is public reading interest, not a trade magazine. The page chosen for each word is the first Wikipedia hit, which is sometimes the wrong topic. The later period is 2025-09-25 through 2026-09-25. The word list was saved before this score.
In the earlier year the equation lined up with 2.8% of the variation across 264 rows. A one-notch rise in industry attention lined up with a -2.54 percentage point change in the next 20-session return (rough size -2.7).
Kept as a paper research result against the S&P 500. Later period, $250,000 paper: finished at $353,497 (+103,497) across 48 entries. The S&P 500, with a one-basis-point spread, finished at $296,191 (+18.5%).
Climate models earn some skill because they include physical laws and because they are scored on years they were not fit to. This page keeps the second habit. It does not have the first. A clean price history makes false patterns easier to see and easier to believe. The project summary is at /summary. The day-by-day filing blotter is at /desk.