How It Works
Forecasting a sequence of numbers is an older discipline with different rules
Prediction over time developed its own methods and its own way of validating them decades before machine learning arrived, and most of the awkward parts are still exactly where they always were.
By Zoya Rahman4 min read
Order is not a detail, it is the content
A time series is a sequence of measurements taken at intervals: sales by week, temperature by hour, network traffic by minute. Shuffle the rows of an ordinary dataset and nothing is lost. Shuffle a time series and everything is. The information lives in the arrangement quite as much as in the values.
That single property invalidates a large part of the standard toolkit. Techniques assuming each row is an independent draw from some fixed distribution are not merely suboptimal here; their assumptions are plainly false, because each observation is related to the ones next to it in a way the model has to account for.
What makes the problem feel harder than it is: the data usually looks trivial. One column of numbers. The difficulty is entirely in the structure hiding inside that column, and the structure is frequently not stable across the length of the record.
The classical methods stated their assumptions out loud
The traditional decomposition breaks a series into a slow-moving trend, a repeating seasonal pattern, and whatever is left over. Each part is modelled separately and then recombined. It is transparent to the point of being teachable on paper, which is not a small advantage when a forecast has to be defended to somebody.
Exponential smoothing weights recent observations more heavily than old ones, with the rate of forgetting as an explicit parameter. The autoregressive family models each value as a combination of previous values plus a shock. Both are decades old, both remain in daily production use, and neither is embarrassed about it.
The important thing about these methods is not their age but their explicitness. They say what they assume — that seasonality repeats at a fixed period, that the noise behaves in a particular way — and when a forecast goes wrong you can usually identify which assumption broke.
Validation has to respect the direction of time
The standard way of testing a model is to hold out a random portion of the data. Do that with a time series and you have trained on Thursday in order to predict Wednesday, which produces excellent scores and a worthless model. Most serious mistakes in applied forecasting begin somewhere near here.
The correct procedure trains on a prefix and tests on what follows, repeatedly, rolling the boundary forward through the record. It is more work, it yields fewer test points, and the resulting numbers look worse — which is the point, because they are the numbers describing what will actually happen next.
Leakage in this setting is subtle. A feature computed as an average over the whole record carries information from the future back into the past. So does any preprocessing step fitted before the split. These errors are easy to make, hard to spot, and they always flatter the result.
Learned methods took longer to win here than elsewhere
Neural networks swept image and speech work relatively quickly. Forecasting resisted, and for a long time the open competitions in the field kept finding that simple statistical methods, or averages of several of them, were difficult for elaborate models to beat across collections of real business series.
The likely reasons are structural rather than mysterious. Individual series are short — a few hundred points, where a vision model has millions of examples — and they are noisy, and much of what drives them is not in the data at all. A model with enormous capacity and a few hundred observations will fit the noise.
What changed the picture was training one model across many related series at once, so that patterns learned from thousands of products or sensors transfer to each individual one. Results from more recent competitions and from practice have been considerably stronger, and the field is no longer unanimous that classical methods are the safe default.
A forecast without a range is only half of one
A single predicted number is almost never the useful output. What a decision needs is a range with a stated likelihood attached: an interval within which the value is expected to fall most of the time. Producing those honestly is harder than producing the central estimate, and they are frequently omitted for exactly that reason.
Intervals from most methods are known to come out too narrow, because they account for the noise the model measured and not for the possibility that the model is wrong about the structure. Real series break their own patterns — a shop closes, a policy changes, a sensor is replaced — and no interval derived from history covers that.
Which leads to the standing disagreement about how far ahead forecasting is worth attempting. Short horizons are largely a technical problem. Long ones depend on conditions the record cannot contain, and a confident long-range number derived from a clean-looking curve is a claim about stability rather than a measurement of anything.
Common questions
Why can a forecast be accurate for months and then fail suddenly?
Because it extrapolates structure found in past data, and that structure holds only while the underlying conditions do. A change in behaviour, supply or policy creates a regime the record never contained. The model has no way of detecting this except by being wrong afterwards, which is not much use in advance.
Is a longer history always better?
Not automatically. More data helps estimate stable patterns, and it also drags in periods when the series behaved differently for reasons that no longer apply. Deciding how much history is relevant is a judgement about the process being measured rather than something the data can settle on its own.
Does adding external variables improve a forecast?
It can, but it introduces a new problem: to use a variable for future dates you must forecast that variable too, and its error compounds into yours. Variables known in advance, such as calendar effects or scheduled events, are far more useful than ones that would themselves need predicting.
Deputy editor, AI Worth Knowing
Zoya joined to cover how it works, in the world, limits & risks and stayed for the awkward questions and would rather show the working than assert the conclusion.
