From now to one hour on
Two forecasts are made for each topic each time it is recomputed: which way the next hour is likely to go, and the chance that it passes 500 posts.

01
Only topics busy enough to call
A topic with fewer than 20 posts in the last hour gets no direction, and the screen says so with the number.
02
The direction
The next hour against the last: rising at one and a half times the last hour’s posts or more, falling at two thirds or fewer, about the same between, by default. Shown as “likely rising”, “likely about the same” or “likely falling”.
03
The expected level
The dashed line is the last hour’s level times the expected change: one and a half for rising, one for about the same, two thirds for falling, weighted by how likely each is.
04
The 500 line
A separate forecast with a fixed meaning: the chance of 500 posts or more inside the next 60 minutes. It is shown on its own line and never as a direction.
05
Each watched keyword too
A keyword gets its own forecast over all its posts together, by the same model and the same rules, kept apart from the topic forecasts.
What goes in
The forecaster reads each topic’s activity minute by minute. Training and live forecasts compute these with one function.
01
Growth
How fast posts are growing over the last 5, 15 and 30 minutes, and whether that growth is speeding up or slowing down.
02
Volume
Posts in the last 15 and 30 minutes; for the direction, posts in the last hour and the last half hour against the half hour before it.
03
Who is posting
How many different accounts are posting, and the share of posts from accounts with high reach.
04
Spread and engagement
How many platforms carry the topic, and the likes, reposts, replies and quotes per post.
05
Opinion and coordination
The pro and anti shares of stance, and the latest coordination score. When either is missing, the model is told it is missing rather than given a guess.
06
Time of day
In Indian Standard Time.
07
The model
A gradient-boosted tree model for each of the two forecasts, run on the installation’s own machines. A forecast costs nothing per call.
A range and a chance
A forecast is never a single number. The drawing shows what one holds.
The last hour (1) runs up to now (2). The expected level for the next hour is the dashed line, inside its shaded range (3). The 500-posts line (4) belongs to the separate threshold forecast and is not read as a direction.
The bar (5) splits one into three: the chances of rising, about the same and falling. The direction shown is the likeliest of the three; the Forecast screen sets out all three for each topic.
Until the forecaster has passed calibration, these chances are the model’s own figures, marked “not yet checked”. Once it has passed, the forecast carries a probability and a calibrated interval.

Many distinct authors are posting
One of the reason sentences the forecaster writes, here the one for the number of different accounts posting. Each forecast gives the inputs that moved it most, three by default, each as one sentence.
The sentences are fixed. Every input has two, one for when it pushes the forecast up and one for when it holds it back: “Many distinct authors are posting” or “Few distinct authors are posting”, “The topic is spreading across platforms” or “The topic is on few platforms”. Which inputs moved the forecast, and how far, is computed from the model’s trees for that one forecast; the sentence is then looked up, never written by a language model.
On the Forecast screen each reason has a bar beside it, to the right for pushes up and to the left for holds back, its length set by how far.
How a forecast is scored afterwards
Every forecast is stored. When its hour has passed, the actual number of posts in that hour, the actual direction and whether 500 was reached are written onto the same record. Those records are the forecaster’s track record.
A forecast is closed without an outcome when its two hours were not seen whole: when it was made before the collector last started, when its topic merged with another inside the two hours, or when collection was paused or incomplete during them. The reason is kept on the record, and the track record counts each reason.
The track record is set beside the plainest forecast there is, that the next hour will be the same as the last, so the model can be compared with it.
Calibration needs at least three weeks of those records and at least 100 cases in which the forecast outcome happened. Until it has passed, every forecast says “not yet checked”, and a rule written on a forecast’s probability does not fire.
The forecaster’s model has been trained on synthetic history. That is one more reason its figures are marked as the model’s own, and the mark stays until forecasts have been validated on the installation’s own history.
See how a forecast for one of your topics was made, and how it was scored when its hour had passed.
Request a demo/methodology/forecasts · How forecasts are made