Time Series Forecasting

A time series model reads one input: the past values of the series itself, usually with a calendar attached. Fitting it means estimating how much weight recent observations carry against older ones, how far a trend should be extended, and which repeating pattern to project forward. Because everything the model knows comes from the series, a change that happened outside it – a price move, a leaflet promotion, a new store in the catchment area, a delisted substitute – is invisible. The model sees only the resulting bump and attributes it to level, trend or season.

• Exponential smoothing (ETS) – weights recent observations more heavily; handles level, trend and a stable seasonal pattern, and needs little history
• ARIMA – models the autocorrelation of a differenced series; wants a longer and reasonably stable history
• Croston‑type methods – for demand with many zero periods; forecast size and interval separately and return a rate per period, not a prediction for a given day
• Regression on time and calendar variables – the point where a pure time series turns into a causal model, once price, promotion or weather enter as regressors

Time series forecasting in practice

The recurring failure is not arithmetic but a broken assumption. A promotional peak that the model reads as a shift in level keeps the forecast elevated for weeks after the campaign; a store opening nearby looks like a downward trend. None of it is in the series, so no time series method can be expected to find it – the answer is causal variables plus a history cleaned of the events that distorted it, not a heavier model. The same applies to the input itself: pure sales data cannot distinguish “nobody wanted it” from “it was not on the shelf”, so periods with stockouts have to be corrected before any model is fitted, or the model learns censored demand.

Model selection deserves the same scepticism. With a large library of candidate methods and two years of weekly data, the method that wins on in‑sample error is often the one that fitted noise; rolling‑origin evaluation at the granularity and horizon at which orders are actually placed is what separates the two. That horizon matters more than the reported one: where a forecast is evaluated one period ahead but the order covers lead time plus the order cycle, the extrapolation that decides the order quantity was never measured. And whatever the residual error turns out to be, it is safety stock that absorbs it.

Veritico STOCK combines statistical methods, machine learning and neural networks and selects, per item, the method with the smallest expected error from a library of roughly eighty, with parameters tuned by genetic algorithms; history is cleaned of extremes, stockouts and promotional effects before the forecast is computed, so the model is not trained on censored sales. Demand sensing compares the most recent weeks against the long‑run history and adjusts the forecast, and series can be built bottom‑up, top‑down or middle‑out – the last of these being the practical route for new and strongly seasonal items. The forecast then feeds replenishment and allocation.

Related terms

• Demand Forecasting
• Forecast Accuracy
• Seasonality
• Intermittent Demand
• Demand sensing
• New Product Forecasting