Forecasting

Forecasting is the practice of projecting a future value from what you already know — past observations, current conditions, and stated assumptions. A demand planner forecasts next quarter's unit sales; a finance team forecasts revenue; a growth team forecasts site traffic or new customers. The methods range from simple to elaborate. A naive forecast repeats the last observed value. A moving average smooths recent history. Exponential smoothing weights recent data more heavily and can absorb trend and seasonality. Regression relates the outcome to drivers such as price or ad spend. Time-series models like ARIMA capture momentum and cycles from the series itself. Judgmental forecasting adds human expertise where data is thin. Whatever the tool, a forecast is an estimate with uncertainty around it, not a fact about the future.

Forecasting earns its keep because almost every plan depends on a number that has not happened yet. Order too little inventory against a demand forecast and you stock out; order too much and cash sits on shelves. Staffing, budgets, cash flow, and capacity all rest on projected figures. A good forecast narrows the range of surprises and lets you commit resources with less risk. The danger is treating the single headline number as certain. Every serious forecast should carry a sense of its own error — a range, a confidence interval, or at least a track record of how far off past forecasts ran. Read that way, forecasting is less about being exactly right and more about being usefully close, and knowing how close you are likely to be.

You cannot judge a forecast the moment you make it — only later, against what actually happened. That comparison is forecast error, and it is measured, not guessed. Mean absolute error averages the size of the misses in the original units. Mean absolute percentage error expresses the miss as a percent, which travels across products of different sizes. Root mean squared error punishes large misses more sharply, so it is useful when big errors hurt most. Bias tracks whether a forecast runs persistently high or low, which is different from being noisy. A forecast can be unbiased on average yet swing wildly, or steady yet consistently too optimistic. Picking the right error metric depends on what a miss costs you, and reporting error is what keeps forecasting honest.

Forecasting is often confused with two neighbors it is not. A forecast is not a target. A target is what you want to happen and will work to cause; a forecast is your best estimate of what will happen given current plans. Confuse the two and you produce forecasts that are really wish lists, and the error metrics quietly rot. A forecast is also not a plan. The plan is the set of decisions — how much to order, hire, or spend — that you make partly in light of the forecast. Keeping the estimate separate from the goal and from the decision is what lets you learn: when actuals arrive, you can see whether the forecast was off, whether the plan was wrong, or both, instead of blaming one for the other.

Good forecasting starts by matching the method to the data and the stakes. Short, stable series may need nothing more than smoothing; series with strong seasonality and trend call for models that separate those components; outcomes driven by known levers may suit regression. Whatever you choose, test it honestly: hold out recent history, forecast it as if it were unknown, and measure the error against what really happened — this is backtesting, and it beats trusting a model that merely fits the past well. Combining several reasonable forecasts often beats betting on one. Update as new data lands, capture seasonality and known events, and write down your assumptions so a forecast that misses can be diagnosed rather than merely regretted. Above all, publish the uncertainty alongside the number.

The failures are predictable. Confusing a forecast with a target turns estimation into advocacy. Chasing a tiny in-sample error by overfitting produces a model that shines on the past and fails on the future. Ignoring seasonality or a known one-off — a promotion, a holiday, a launch — bakes an obvious miss into the number. Reporting a single figure with no range invites false confidence, and never checking forecasts against actuals means the same errors repeat forever. A subtler failure is anchoring to one favored method out of habit, when a blend — a smoothing model for momentum, a regression on known drivers, and a measured dose of expert judgment — would have been steadier and easier to defend. The discipline is the opposite: choose a method suited to the data, backtest it, quantify and communicate uncertainty, keep the forecast separate from the plan and the goal, and treat every realized error as feedback that sharpens the next forecast.

Forecasting projects future values from historical data and assumptions, and is judged by forecast error rather than by any claim of certainty.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where forecasting is a core concern: