About GTN Alpha
Every day, for every covered stock, the Kernel reads what was published, weighs who said it by how often they have been right, and posts a decision: BUY, HOLD or SELL, at three horizons, with a confidence from 0 to 100 and the evidence that drove it. Then it grades yesterday’s decisions against the price and learns.
This page says exactly how — what is collected, what is extracted, how claims are graded, how weights are learned, what a decision means, and why the baselines are never hidden. Nothing in the loop is hand-set, and nothing is a black box.
Financial news as it is published, captured once and kept as it was: the Interactive Brokers news wires (Dow Jones, Morningstar bulls & bears, Campus), The Motley Fool, and IBKR fundamentals — daily prices, consensus estimates, the earnings calendar. Captures are durable and never restated with hindsight; a story that appears in forty outlets is one story, and the copies are linked to it.
Each article is distilled into attributed, testable claims: who asserted what, about which company, in which direction, on which metric, by when, under what conditions, and whether they were bullish, bearish or neutral. Alongside them, factor observations — a call on an industry input such as memory availability or rates — are propagated to the tickers exposed to it through a curated exposure map, so a supplier’s tightening becomes a signed input for its buyers.
The extraction is a language model reading a common capture format for every source; the model never branches per outlet. Its accuracy is not independently measured — the owner chose to go with it and let the scorecard be the judge.
Against realized price only. When a claim’s horizon ends, the return from the close on the day it was made to the close at its horizon is measured. A bullish call is correct if the price rose through the dead band, incorrect if it fell through it, and inconclusive in between; bearish calls mirror that; a neutral call is correct only if the price stayed inside the band. The band widens with the horizon — ±2% short, ±4% mid, ±6% long — so noise is not mistaken for foresight.
A revenue, margin or earnings claim is graded the same way: the owner accepts that price is the target, and the claim is judged by whether the market moved as it said. A syndicated story earns its author one grade, not one per copy. Factor observations are graded against every exposed ticker, with the sign the exposure map gives.
No hand-set weights. Every weight comes from the grading ledger and is recomputed daily: for each
author, outlet, author class, factor, horizon and metric, a smoothed hit rate
2 · (correct + 2) / (correct + incorrect + 4). An unmeasured voice sits at exactly 1.00; a record has to be
earned before it can move the weight, and thin records are shrunk toward the middle so nobody dominates on three
lucky calls. Management talking its own book is learned like everyone else, not zeroed by decree.
Weights are point-in-time: the weight used on a day is built only from grades that had matured by that day. Replaying the past never borrows from the future.
For each ticker and horizon, the live claims and factor inputs are a weighted vote: bullish +1, bearish −1, neutral 0, each multiplied by the product of its learned weights. The vote is a score from −1 to +1; confidence is that score’s strength scaled by how much weighted evidence stands behind it, so one loud voice cannot reach 100. BUY when the score clears the buy cut, SELL when it falls below the sell cut, HOLD between.
The cuts are learned too: each day they are chosen to maximize hits minus misses over past decisions that have already matured at a fixed lead time — never fitted to the day itself. One dated snapshot is written per day and can be replayed as of any date. The top drivers are always shown with the decision; a decision is never a black box.
Every day the scorecard reports the weighted lane’s hit rate and mean return next to two baselines that cannot be switched off: unweighted consensus — the same evidence with every voice counted once and equally — and always-BUY — buying every ticker every day. The first says whether learning helps; the second says whether the whole exercise beats doing nothing clever.
Two months of backtests before this lane found no significant signal. That is the prior the system inherits, and the baselines are what keep the claim testable. If the weighted line does not sit above both, the scorecard says so, in the open.
Cadence
Grade → learn → decide → mature → score, once a day. The family executes on a daily or weekly cadence. Every day’s decisions are kept as they stood, so “what did it say on the 3rd?” always has an answer.
Echoes
When 47 newsletters repeat the same story, it isn’t confirmation — it’s a louder version of one opinion. Syndicated copies are collapsed before the vote and before the grade.
Evidence
The direct-claim tilt and the factor lane stay visible as evidence on each ticker page. The decision nets them into an action; the evidence display does not.
Grade against the price. Learn the weights. Show the baselines.
— the three owner rulings the lane is built on