A global forecasting model trained on every show in the portfolio at once

Eight shows, eight forecasting models, eight sets of coefficients, and a ninth show acquired in March that has no model at all because it has no history with you. Every year the same conversation happens about whether the new one can be forecast yet, and every year the answer is that it needs a couple of editions first.

A global forecasting model removes that conversation. One function is fitted across every series in the portfolio at once, show identity goes in as a feature the model reads, and a new edition inherits the shared pattern from the first week of its campaign.

What global changes about the fit

The distinction is about how many functions you estimate. The local approach treats each show as a separate problem and fits a separate model to each. The global approach fits one model to all of them.

Montero-Manso and Hyndman set the two side by side in the International Journal of Forecasting in 2021 and reached a conclusion that goes further than most teams expect. Their first stated contribution is that "Global methods are not more restrictive than local methods, both can produce the same forecasts without any assumptions about similarity of the series." A global fit does not require your shows to look alike. It requires the model to be given enough to tell them apart.

Their second contribution is the one with operational consequences. The complexity of local methods grows with the size of the set, they write, while it remains constant for global methods. Eight shows means eight times the parameters under the local approach and roughly the same number under the global one, which is why they conclude that in large datasets a global algorithm can afford to be quite complex and still generalise better.

The parameter arithmetic, on eight shows

Take weekly cumulative registrations, forty weeks per campaign, eight shows, five editions each. Suppose the model is autoregressive on the weekly pacing series.

Fit locally with three lags. Each show needs three lag coefficients plus an intercept, which is four coefficients per show, and across eight shows that is 32 coefficients to estimate. Each of those coefficients is supported only by that show's own 200 weekly observations.

Now spend the same 32 coefficients globally. Twenty four lag coefficients shared across every show, plus one intercept per show, is 24 plus 8, which is 32. Identical budget. The global version has 24 weeks of memory where the local version had 3, and every one of those 24 coefficients is estimated from 1,600 weekly observations instead of 200.

This is exactly the trade Montero-Manso and Hyndman describe when they say that for autoregressive models the bound implies global models can have much larger memory than local methods. In their empirical work they report that global linear models can provide competitive accuracy with two orders of magnitude fewer parameters than local methods, a factor of a hundred rather than a rounding difference.

For a show forecast, 24 weeks of memory matters more than it sounds. Three weeks of lag can see a plateau. Twenty four weeks can see where the campaign sat when the early bird deadline passed, which is usually the thing that determines whether the current plateau is a problem.

Hold the lag length fixed instead and the same arithmetic buys precision. A global model with three shared lags plus eight show intercepts needs 11 coefficients against 1,600 weekly rows, which is 145 observations per coefficient. The local version needs 32 coefficients against those same 1,600 rows, which is 50. You can spend that difference on memory, on features, or on nothing at all and simply keep the tighter estimates, which is the option nobody ever picks and the one most defensible on a portfolio this size.

Why does a global model handle a brand new show better?

Because the shared coefficients are already estimated, and the only thing the new show contributes is its own intercept.

A ninth show joins the portfolio. Under the local approach it has no coefficients and no way to get any until it has run a few editions. Under the global approach the 24 lag coefficients apply to it on day one, and the show-specific intercept can be estimated from the first several weeks of its own registration file, or set to the portfolio average until there is enough to move it.

That does not make the first forecast good. It makes it available, and available with a sensible shape rather than a straight line. A genuine first edition with no registration history at all is a different problem again, and cold start forecasting is O35's.

The same logic covers the awkward middle cases a portfolio produces constantly. A show that skipped a year. A show that changed name. A show that ran a half-size edition during a venue refurbishment. Each of those breaks a local model and each is absorbed by a global one as a show with less usable data than its neighbours.

Keeping shows distinguishable inside one fit

The failure mode of a global model is homogenisation. If every show gets the same coefficients and only the intercept varies, then a show whose campaign genuinely peaks eight weeks earlier than the rest gets the average peak, and its forecast is wrong in a consistent direction that nobody notices because it is wrong in the same way every year.

The fix is features, and they have to be features that actually separate the shows. Useful ones on a portfolio of trade shows are the vertical, the hall capacity in net square metres, the number of editions the show has run, the position of the show in the calendar, the number of weeks between registration opening and doors, and whether the current edition changed venue. Each one gives the model a legitimate reason to bend the shared curve for a particular series.

There is a test for whether you have enough of them. Fit the global model, then compute the residual bias per show, which is the mean residual for each series. If one show's residuals average well away from zero across its whole campaign, that show has a characteristic the features do not describe, and the model is quietly reallocating its error to everyone else. Either add the feature that explains it or fit that show separately and say so.

The alternative construction, where the per-show deviation is an explicit random effect shrunk toward zero by a weight you can read off, is the multilevel version, and borrowing strength across shows covers it in O14. The two approaches overlap heavily in what they compute. They differ in how visible the per-show adjustment is, and the multilevel version is easier to explain to a show director who wants to know why the model disagrees with them.

Does a portfolio of eight series count as a large dataset?

Not by the standards of the literature, and this is worth being honest about internally before someone builds a case on it.

Salinas, Flunkert and Gasthaus described the global approach in its modern form in 2017, training an autoregressive recurrent network on what they call a large number of related time series, and reported accuracy improvements of around 15 per cent against the methods they compared with. Their datasets run to thousands of series. Eight shows with five editions each is 1,600 weekly rows.

The generalisation argument still applies at that size, because it is an argument about the ratio of parameters to observations rather than about absolute scale. What changes is the size of the prize. With eight series the gain comes from the shared weekly pacing pattern, which is real and worth having. It does not come from the model discovering structure across a thousand series, because you do not have a thousand series.

The practical reading is to keep the global model small. A global linear or gradient-boosted model on a few well-chosen features will do the work. A deep network on 1,600 rows is the same overfitting problem that every candidate model fitting five points describes in O12, moved to a bigger table.

Where this stops

A global model makes one large assumption: that the relationship between the features and the outcome is the same for every show. When that fails, it fails invisibly.

Suppose seven of your shows are exhibitor-led events where registrations arrive steadily and one is a conference where two thirds of the file lands after the programme is published. The programme publication date is a feature that means something for one show and nothing for the other seven. A global model will fit a small average effect, which understates it for the conference and adds noise to the rest. Interactions between show identity and the key features are the repair, and each interaction spends parameters you were trying to save.

The second limit is auditability. Eight local models are eight things a show director can be shown. One global model is a single object whose behaviour for any individual show has to be extracted rather than read, and in a business where the forecast gets argued about in a room, that cost is real. Stacking the same data as show-year rows and running an estimator whose coefficients everyone can read is the more transparent option, and pooled panel regression sets that out in O16.

Take your weekly registration snapshots for every show and every edition you hold, stack them into one table with columns for show, edition, days to open and cumulative registrations, and count the rows. That number, against the number of parameters your current per-show models consume in total, is the whole argument for or against going global, and it takes an hour. What a platform needs to hold to make that table possible sits with the forecasting methods behind it.