How SAC's time-series forecasting predicts future trends from historical data — the key concepts (trend, periodicity, fluctuations, residuals) and a real retail use case.
A time series is a sequence of data points arranged chronologically, each recorded at regular intervals. Companies track metrics like revenue, expenses and inventory at consistent intervals to identify trends and patterns — enabling informed predictions about future performance.
Understanding signals
Effective modelling pairs target variables with corresponding dates — the 'signal'. SAC's Smart Predict examines this signal and can incorporate additional variables as influencers. Signals contain four components: Trend (the overall direction), Periodicity (seasonal or cyclical patterns), Fluctuations (how current values depend on past values), and Residuals (random, unpredictable noise).
Model creation
Smart Predict trains multiple trial versions from the training dataset, selects the best performer, evaluates it against validation data, and generates the final model. Key considerations: the scale of predictions (historical intervals determine granularity — monthly data can't forecast daily outcomes), appropriate aggregation (smoothing noise without obscuring information), and clean data sorting (each time unit corresponding to one target value).
Use case: retail sales forecasting
Predicting sales across thousands of product families at stores in Ecuador, using training data of dates, store numbers, product families, promotional status and historical sales — with an 'entity' column identifying each forecasting subject (e.g. store-product combinations) so the model generates separate forecasts for each.
Conclusion
Time-series forecasting in SAC enables organisations to predict future trends grounded in historical patterns — and understanding trend, seasonality and fluctuations supports better decisions and strategic optimisation.
