Direction agreement is the share of stocks where our forecast and a completely independent third-party model point the same way — both up, or both down. It is a corroboration check: a second opinion from a model that has no idea ours exists.
There is an obvious problem with any company grading its own homework. We built our forecasting model, we chose how to test it, and we publish the results. Even with the best intentions, that is a closed loop.
One way to open the loop is to bring in a model we did not build, did not train, and cannot tune — run it on the same stocks over the same periods, and see whether it reaches the same conclusions. Where it agrees with us, that is a small piece of independent corroboration. Where it disagrees, that is a flag worth taking seriously.
Chronos-2 is a time-series foundation model published by Amazon. A "foundation model" here means the same thing it means for language models: it was pre-trained on an enormous, diverse collection of time-series data, and it can then forecast a series it has never seen before without being retrained on it. That last property is called zero-shot forecasting.
The relevant point for you is what it is NOT. It is not tuned to our data, our tickers, or our preferences. We cannot make it agree with us. That independence is the entire reason it is worth showing — a second opinion is only worth something if the second opinion is genuinely free to disagree.
Which of the two models draws the more honest confidence band is not settled here in prose — it is measured, separately at each forecast start date, and published on the calibration page whichever way it comes out. We do not show Chronos-2 because it loses to us. We show it because an independent check is more credible than a self-assessment, and a check that could only ever agree with us would be worth nothing.
For each stock in the test basket, both models produce a forecast. We ask a single yes/no question of each: relative to the starting price, does this forecast point up, or down? Direction agreement is simply the share of stocks where both models gave the same answer.
Note what is deliberately thrown away. The size of the move does not matter; the confidence does not matter; whether either model was actually RIGHT does not matter. It is a pure agreement count between two opinions, which is what makes it a corroboration signal rather than a performance metric.
The asymmetry is the useful part. Agreement should not make you bolder, but disagreement can reasonably make you more cautious.
When both models point the same way, you have learned relatively little: the easy cases are exactly where independent models converge, so agreement is cheapest precisely when it is least needed. When they point opposite ways, you have learned something real — this stock sits in genuinely contested territory, where two different methods reading the same history reach opposite conclusions. That is a good moment to want a wider margin, a smaller position, or no position.
We tested whether to formalise this — gating our conviction score on agreement with Chronos-2 — and it produced no measurable improvement in directional accuracy. So we do not do it. Agreement is information for you, never a hidden input to our signals, and saying so plainly is more useful than quietly adding a feature that did not work.
The agreement figure is published on the calibration page, beside the confidence-band comparison and measured the same way: separately at each forecast start date, shown as the range across those dates rather than as one number. It had dropped off the page while it was only measured from a single start date — a single start date tells you about that date rather than about either model, which is the same reason every figure there is shown start date by start date.
The core trap is treating agreement as confirmation. Two models trained on the same historical market, both looking at the same price history, are not truly independent observers of the future — they share the past. When they agree, some of that agreement reflects a shared blind spot, not converging evidence. In a genuine change in market conditions, models tend to be wrong together.
The second trap is treating a directional call as the whole forecast. "Up" is not a plan. Two models agreeing on "up" tells you nothing about how far, how likely, over what horizon, or what happens if it goes the other way. The band, the entry level and the invalidation level are where the actionable content lives.
Third, do not confuse this with the directional hit rate, which measures how often a model got the direction RIGHT against what actually happened. Direction agreement never checks reality at all — it only checks whether two models said the same thing. They sound similar and answer completely different questions.
The share of tested stocks where our forecast and an independent third-party model (Amazon Chronos-2) point the same way relative to the starting price — both up, or both down.
A time-series foundation model published by Amazon: pre-trained on a large, diverse collection of time series, and able to forecast a new series without being retrained on it. We do not train or tune it, which is exactly what makes it useful as an independent second opinion.
That is not a question we answer in the abstract. What we measure is whether each model's 90% band actually contained the price 90% of the time, separately at every forecast start date we test — and the answer moves a lot depending on the start date. The calibration page publishes every one of those comparisons, including the horizons where Chronos-2 came out ahead. We publish it because an independent cross-check is more credible than grading our own homework, not because we win it.
No. Agreement is not accuracy — two models can agree and both be wrong, and they agree most readily on the easy cases. Disagreement is the more informative signal, and it is a reason to lean toward caution, never a reason to buy.
Browse all S&P 500 tickers to see this metric applied to individual companies.
Educational research only — not investment advice.