Organizations worldwide use statistical, time series forecast methods to predict myriad conditions, such as the weather, product demand, emergency service requirements, and more. Time series forecasts are the basis of predictive analytics, and of how you plan, meet your requirements, and set strategies. For example, hotel operators use time series forecasts to predict the number of guests over a given period. This is how they map out staffing and hotel inventory requirements, as well as revenues, expenses, and where and when to expend cash-flows.
At the most basic level, a time series forecast starts with capturing and analyzing historical date, such as sales transactions, visitor volumes, or equipment breakdowns. Then statistical models are applied to these data to identify historical trends and patterns that you can extrapolate into the future.
A trend might be the steady uptick of sales over the course of several years. Will it continue? You don’t know unless you apply statistical models to identify correlations across different data sets. Using regression analyses, you can model relationships between independent and dependent variables, as well as sensitivities to change that help make your assumptions about cause and effect more reliable. The same is true for identifying and understanding patterns, such as the seasonal drop-off in beach-front visitors during winter months, or the surge in agro-chemical sales during the pre-planting season. In such cases, time of year is a strong determinate of transactional activity.
Put more broadly, commerce is a function of demand drivers that occur as a matter of time, location, and action. These include seasons, climate, weather events, holidays, cultural events, consumer sentiments, fashion fads, viral tweets, and more.
Following are more complex time series techniques for extrapolating historical data into the future.
Exponential smoothing (ETS) models
Exponential smoothing is a time series forecasting technique that applies a weighted average of past values to forecast the future. Unlike simple moving average forecasts, which weight data equally, exponential Smoothing gives higher weights to more recent data, which is useful for picking up more recent trends and patterns. As such, older data in the selected time series influence the forecast, but with decreasing strength as those values age.
In addition to picking up recent trends and patterns, exponential smoothing techniques are also used to smooth out steep fluctuations in forecasts that might result from outlier events or other forms of random noise.
Exponential smoothing forecast methods include:
- Simple exponential smoothing
- Holt’s linear method
- Winters’ multiplicative season
- Winters’ additive season
- Croston’s intermittent demand
- ARIMA (Box-Jenkins)
Decomposition forecasting methods are designed to decompose, or separate, historical data into various component factors that affect the forecast. Decomposition looks beyond the simple trend line to also calculate the effects of annual seasonality, day-of-the-week patterns, holiday effects, a business cycle, and other factors that, taken into account, improve the accuracy of the forecast.
Decomposition methods are good at daily forecasting, as well as at predicting pattern-based fluctuations around underlying trends, such as holiday retail sales that affect the fourth quarter.
The more components that are factored, the more difficult the computation and analysis, which can quickly outstrip the limitations of manual computing in spreadsheets.