What accuracy means and how it is measured
Forecast accuracy is simply the gap between what a model predicted and what actually happened, expressed as an error rate. The smaller the error, the more closely the forecast tracked reality. It is measured on holdout data - periods the model never saw while it was learning - so that the score reflects genuine predictive skill rather than memorised history. A model that fits the past perfectly but fails on unseen data is not accurate; it is overfitted.
The metric matters as much as the number. ElenIQ reports WMAPE (weighted mean absolute percentage error) as its primary measure because it weights each error by the volume behind it, so a busy channel is not flattered by a quiet one and a tiny denominator cannot blow the score out of proportion. sMAPE sits alongside it as a secondary check, and every model is compared against a naive baseline - the error you would get by simply assuming next period looks like the last. Beating that baseline is the bar a forecast has to clear before it earns any trust at all.
Why no forecast is perfect - and how accuracy informs confidence
A forecast models behaviour that has not happened yet, and markets carry noise that no model can remove. Seasonality shifts, competitors change tactics, creative fatigues, and demand moves with the wider economy. A model can only learn from the patterns present in past data, so a residual amount of error is irreducible. The aim is never zero error - it is an error small enough that you can plan against the forecast with confidence.
This is exactly why accuracy and uncertainty are reported together. A model’s historic error feeds directly into the width of its prediction bounds: high error widens the range of plausible outcomes, while low error tightens it. Reading the headline number next to its confidence interval gives you both the central estimate and a realistic sense of how far the result could swing - which is what you actually need when committing budget. Accuracy is the foundation of disciplined paid media forecasting: without it, a forecast is just a confident-looking guess.
How to use accuracy to decide whether to trust a forecast
Accuracy is the dial that tells you how hard to lean on a forecast. Consider a planner deciding whether to move £50,000 into a channel. A model that scores a 9% WMAPE on holdout data has tracked reality closely enough that its forecast is trustworthy to act on - the likely outcome is well understood and the bounds are tight. A model scoring a 45% WMAPE, by contrast, is little better than guessing; its predictions should be treated with caution, the £50,000 decision should not rest on them alone, and the sensible response is to gather more data or test at smaller scale first.
Crucially, low accuracy does not mean you do nothing - it means you size the bet to the evidence. A weak forecast still narrows the field; you simply commit less, hedge across options, and revisit as actuals arrive. This is where accuracy meets scenario planning: when the central estimate is uncertain, you model a range of budget outcomes and choose the allocation that holds up across them rather than betting everything on a single point prediction.
Why forecast accuracy matters for budget allocation
Budget is allocated at the margin, before the spend is committed, and every allocation rests on a forecast of what the next pound will return. If you cannot judge how reliable that forecast is, you cannot judge the risk of the decision. Accuracy turns a prediction into something you can weigh: reliable forecasts justify decisive moves, while shaky ones call for smaller, staged commitments. Used well, it stops teams over-investing on the strength of numbers that were never trustworthy.
ElenIQ reports WMAPE on every model it trains, alongside the baseline comparison, so the accuracy of a forecast is visible at the point of decision rather than discovered after the budget is spent. That transparency is the backbone of paid media forecasting for agencies, where client budgets demand a defensible basis for every recommendation. To pressure-test an allocation against a forecast you can trust, model the trade-offs directly in the budget allocation simulator.
Related terms
- confidence interval - the range a forecast is likely to fall within, set by its historic error.
- paid media forecasting - predicting channel performance before budget is committed.
- scenario planning - modelling a range of budget outcomes when the central estimate is uncertain.
Frequently asked questions
What is forecast accuracy?
Forecast accuracy measures how close a model’s predicted outcomes are to the results that actually occurred. It is usually expressed as an error metric such as WMAPE or sMAPE, where a smaller error means a more accurate forecast. In media planning it tells you how much weight to put on a forecast before you commit budget against it.
How is forecast accuracy measured?
It is measured on holdout data the model never saw during training, by comparing predictions to actuals. ElenIQ reports WMAPE (weighted mean absolute percentage error) as its primary metric because it weights errors by volume and is not distorted by small denominators, with sMAPE as a secondary check. Both are compared against a naive baseline to confirm the model genuinely beats a simple guess.
Why is no forecast ever perfect?
Forecasts model behaviour that has not happened yet, and markets carry irreducible noise - seasonality shifts, competitor moves, creative fatigue, and economic conditions all change. A model can only learn from the patterns present in past data, so some error is unavoidable. The goal is not zero error but error small enough to plan on with confidence.
What WMAPE counts as a good forecast?
It depends on the channel and the volatility of the data, but as a rule a model around 9% WMAPE on holdout is reliable enough to act on, while one near 45% is little better than guessing and should be treated with caution. Lower-volume or highly seasonal channels naturally carry higher error, so always read accuracy alongside the baseline comparison rather than against a fixed threshold.
How does forecast accuracy relate to confidence intervals?
Accuracy and uncertainty are two views of the same thing. A higher historic error implies wider prediction bounds, so a low-accuracy forecast produces a wide confidence interval and a high-accuracy forecast produces a tight one. Reading the two together tells you both the central estimate and how much it might move.