A forecast without a range is not a forecast
Someone asks what next month looks like and you give them a number. They write it down, plan around it, and it is wrong — as every forecast is. The number was never the problem. Giving it without a range was.
Why the interval widens. Predicting one step ahead, you know almost everything relevant: where the series is now, which way it is heading. Ten steps ahead, each step's error feeds the next, so uncertainty compounds. A forecast chart with a band of constant width is claiming the tenth month is as knowable as the first. It is not, and the widening cone is the honest picture.
Exponential smoothing. Holt-Winters carries three things forward: the current level, the current trend, and a repeating seasonal shape. Each is updated by a weighted blend of what just happened and what was expected — the weights are α, β, γ, and SenSight fits them to your data. It is simple and, on short business series, very hard to beat.
ARIMA. Instead of tracking components, ARIMA models the dependence between a value and its recent past, after differencing away any trend. It is more flexible and it needs more data. SenSight searches a small grid of orders and picks by AIC — but the order it picks is a choice, not a discovery. When the runner-up is nearly as good, that is said out loud, because reading a specific (p, d, q) as a fact about the machine is a mistake.
⚠️ The number that actually matters: the back-test. Every model fits the data it was shown — that is what fitting means. AIC and R² measure exactly that and are always flattering. So SenSight hides the last stretch of your series, fits on the rest, forecasts the hidden part, and reports the real error. That is the only number that answers "how wrong is this likely to be".
And the comparison that keeps you honest. The naive forecast — next month equals this month, or equals the same month last year — is embarrassingly good. If your model cannot beat it, the model is not earning its complexity, and you should either fix it or use the naive one and spend the effort somewhere it matters. SenSight reports both, every time, so the comparison is never optional.
Why the interval widens. Predicting one step ahead, you know almost everything relevant: where the series is now, which way it is heading. Ten steps ahead, each step's error feeds the next, so uncertainty compounds. A forecast chart with a band of constant width is claiming the tenth month is as knowable as the first. It is not, and the widening cone is the honest picture.
Exponential smoothing. Holt-Winters carries three things forward: the current level, the current trend, and a repeating seasonal shape. Each is updated by a weighted blend of what just happened and what was expected — the weights are α, β, γ, and SenSight fits them to your data. It is simple and, on short business series, very hard to beat.
ARIMA. Instead of tracking components, ARIMA models the dependence between a value and its recent past, after differencing away any trend. It is more flexible and it needs more data. SenSight searches a small grid of orders and picks by AIC — but the order it picks is a choice, not a discovery. When the runner-up is nearly as good, that is said out loud, because reading a specific (p, d, q) as a fact about the machine is a mistake.
⚠️ The number that actually matters: the back-test. Every model fits the data it was shown — that is what fitting means. AIC and R² measure exactly that and are always flattering. So SenSight hides the last stretch of your series, fits on the rest, forecasts the hidden part, and reports the real error. That is the only number that answers "how wrong is this likely to be".
And the comparison that keeps you honest. The naive forecast — next month equals this month, or equals the same month last year — is embarrassingly good. If your model cannot beat it, the model is not earning its complexity, and you should either fix it or use the naive one and spend the effort somewhere it matters. SenSight reports both, every time, so the comparison is never optional.
Try it in the app
Try: Time Series tab, Forecast. Run it once, read the back-test line first, and only then look at the numbers — then switch method and see whether the back-test actually improves.