The live nowcast

The horizon is chosen by the people, not the model. A host can act on a cue within a minute: pick up the serum, say the price, ask for the order. A cue takes about two minutes to land in the ledger, because the viewer who hears it has to tap the card, read the page and pay, and session 2 found the stream itself carries a delay before the words even arrive. So one minute is too short to see anything change, and two is the earliest a change shows. Five is the shortest horizon a producer can plan against: long enough that a cue given now is fully inside it, short enough that the room at the end of it is still recognisably the room at the start.

Go longer and the forecast gets easier to hit and useless to act on. A thirty-minute forecast is mostly the room's slow drift, which the producer can see with her own eyes. The five-minute number is the one she cannot see: whether the next five will be better or worse than the last five, and by roughly how much.

• Nowcast - a forecast of the immediate future from the immediate past, made while the thing is still happening. Distinguished from a forecast by the horizon (minutes, not weeks) and by the fact that the same model runs again every minute with one more row.
• Horizon - how far ahead the number reaches. Five minutes here. The backtest scores the sum over the horizon, not each minute inside it.
• EWMA - exponentially weighted moving average. rate = alpha * new + (1 - alpha) * rate. Small alpha means long memory; the newest minute barely moves the estimate. Large alpha means the estimate chases the newest minute.
• Backtest window - one (stream, minute t) pair at which the nowcaster is run using only rows before t and scored against rows t to t+4. Every window is a fresh forecast; 6,720 of them tonight.
• MAE - mean absolute error, in the unit of the thing forecast. Orders per five-minute window here. It is the average size of the miss, in either direction.
• Base rate - how often a rule fires when nothing is done about it, and how often a firing is followed by the thing it was supposed to predict. An alert without one cannot be judged.
• Alert - a rule that turns the nowcast into a yes or no. It is the only part of tonight a producer ever sees, so it carries the most design.

Orders in a minute are the room times the rate at which that room is ordering. Both move, but not equally. Across a five-minute window the per-viewer order rate of a given host in a given stream is nearly a constant with noise on it: one minute has three orders from a room of 800, the next has nine, the one after has four. Smoothing it is the right treatment. The room is different. It rises when the feed sends a burst of arrivals and falls when a segment goes stale, and those moves are real and they persist for minutes. A projection of the room is the right treatment for that.

So the nowcaster smooths the thing that is mostly noise and projects the thing that is mostly signal. The rate is an EWMA of orders[t-1] / room[t-1]. The room projection takes the change over the last five minutes, divides by five for a per-minute slope, and carries it forward five steps with a damping of 0.6 at each step, so a slope of plus twenty viewers a minute contributes twelve to the first projected minute and never compounds into a room that doubles. Each projected minute's room times the smoothed rate is that minute's expected orders; the five of them add up to the nowcast.

That is all of it. There is no seasonality, no exogenous regressor, no state vector. A model a producer can hold in her head is a model she will believe when it disagrees with her, and the backtest is what earns it the right to.

A backtest replays history as if it were live. For every stream and every minute t from 3 to 115, the nowcaster is given rows 1 to t-1 only, asked for the sum of orders in minutes t to t+4, and scored against what actually happened. Minute 3 is the first with enough history for a slope; minute 115 is the last whose five-minute horizon fits inside the 120. That is 112 windows a stream, 6,720 across the 60. The score is the mean absolute error over all of them, in orders per window.

The naive baseline is computed in the same loop, at the same windows: the sum of orders in minutes t-5 to t-1, offered as the forecast for t to t+4. "The next five minutes will look like the last five" is what a producer does in her head, and it is not stupid; it scores 11.97. Any nowcast that cannot beat it has no business in an earpiece. The rule is simple and it is not negotiable: the naive number goes in the same table as the model's number, every time, or the model's number means nothing.

Three alphas, three MAEs: 0.1 gives 9.41, 0.3 gives 10.84, 0.6 gives 11.28. The ordering is monotone and the gap is not small. With alpha 0.6 the rate estimate is mostly the last minute, and the last minute's rate in a room of 800 is a count of three or nine divided by 800, a number with enormous relative noise. The nowcaster then multiplies that noise by five projected rooms. With alpha 0.1 the newest minute gets a tenth of the weight and the estimate is effectively an average over the last ten or so, which is where the true rate lives.

This is a property of the data, not a law. A stream whose host changes what she is doing every two minutes has a rate that genuinely moves, and a slow alpha would lag it. The 60 Lumen Glow streams have segments of a few minutes each and a rate that mostly rides on the host, so the slow memory wins. The sweep in the build-along is there so you re-run it on your own streams rather than inheriting 0.1 as a constant.

Tonight's nowcaster is the bottom rung of a ladder, and it is worth knowing the next two by name so you reach for them when the bottom rung stops beating naive.

• statsforecast 2.1.1 (Apache-2.0): AutoETS, AutoTheta and MSTL on a frame of unique_id, ds, y with freq="min", then sf.forecast(df, h=5, level=[90]) for a prediction interval as well as a point. MSTL takes season_length=[60] or a list if a stream has a repeating segment cycle. One caution: the package has had no release since 2024-09. It is stable and it works; it is not being extended.
• statsmodels 0.15.0 UnobservedComponents with level="local level": the textbook Kalman filter for a true room hidden under noisy per-minute reads. y_t = mu_t + e_t, mu_t = mu_{t-1} + eta_t; the filtered level is the room estimate and res.forecast(5) is the projection. It replaces the slope-and-damping rule with something that learns how noisy the room reads are and how fast the true room drifts.

Both need the same backtest loop and the same naive bar. A forecaster that is more sophisticated and scores 11.5 has lost to the arithmetic. The full method ladder, from naive through ETS and Theta to state space and beyond, with the backtesting discipline that goes with each rung, is a course on its own: learn time series forecasting with phoebe.

The alert has two conditions and both must hold. First, the last three minutes of orders sum to under 70 percent of what the nowcast expected for those three minutes. Three, not one, because one minute under forecast happens constantly in a room where the count is a handful; three in a row is a slide. Seventy percent, not ninety, because a five-order miss on a forty-order forecast is noise and a fifteen-order miss is a room that has stopped buying. Second, the host's last CTA was six or more minutes ago.

This rule, followed exactly, is the "CTA on a dip" policy on session 8's replay bench: when it fires, the host is cued to do a CTA, and over 40 seeded nights of tonight's stream (Jules, Friday 20:00) that policy takes 20,847 dollars against the host's own plan at 18,198, plus 14.6 percent. The bench computes those means live in the browser. The rule you write tonight and the rule the bench runs are the same code.

The second condition is the one people want to delete, and it is the one that makes the rule usable. Without it, the rule fires during CTAs that are already running. A CTA minute has its orders arriving one to two minutes later, so the minute of the ask itself often looks like a shortfall; the rule would then cue a CTA into a CTA, the host would be asked to ask again, and the third time it happened she would take the earpiece out. Six minutes is two for the CTA to land, then a few more of the room converting on it, before it is reasonable to ask a second time.

Before a rule goes in anyone's ear, run it over the 60 historical streams with nothing done about it and count two things. How many times a night it fires on the host's own plan, which is the number of interruptions the producer is signing up for. And of those firings, what share were followed by an actual five-minute shortfall against the nowcast, which is how often the interruption was warranted. The build-along computes both from the history and prints them as a table; they are properties of your streams and your thresholds and they are not the same for a small room as for a large one.

Then move the thresholds and watch both numbers move. Loosen 70 percent to 85 and the rule fires more often on less; tighten six minutes to ten and it fires less often and later. There is no right setting, there is a trade the producer should make with the two numbers in front of her, and a rule shipped without them is a rule whose trade was made by whoever wrote the first draft.

The producer does not hear "fire". She hears a sentence: "orders under forecast, last CTA 8 minutes ago". That is the two conditions read out, and it is deliberately not an instruction. She can see the host, the chat and the product table; the rule can see three columns. She decides whether the cue is a CTA, a price reminder or nothing, because the host is mid-story and the story is working. The rule's job is to make sure she looks, at the right moment, with the reason in her ear.

Two words of the sentence carry the design. "Under forecast" says the shortfall is against what the room was expected to do, not against a number from a different night. "8 minutes ago" says there is room for an ask. When the leader track's session 5 teaches producers to coach live, this is the sentence they practise reacting to.

The nowcaster reads two columns of the minute table, orders and room, and knows nothing else. A change in how the feed is routing traffic shows up in the room only after it has happened, so the projection is late by exactly the amount it matters. A competitor's flash sale at the same hour pulls viewers out of the room for reasons the room curve cannot explain and the rate cannot anticipate. A product going out of stock stops orders dead while the nowcast keeps forecasting them for five more minutes.

None of that is a reason to add columns. It is the reason the horizon stays at five minutes. Every one of those shocks is invisible before it arrives and obvious within two minutes of arriving, and a five-minute forecast re-run every minute is wrong about it for at most a few windows before the new rate and the new slope take over. A thirty-minute forecast would be wrong for thirty. Short horizons are not a limitation the ladder grows out of; they are how a model with two inputs stays honest about the things it cannot see.