Think of Supply Chain Management, the first thing that comes to the mind is Demand Forecasting. Demand or Sales Forecasting is an integral part of the Sales & Operations Planning (S&OP) process of an organization. Demand Forecasting is an area where statistical analytical tools are widely used. There are various statistical models for demand forecasting, which includes Time Series, Causal models or Composite models (which are a combination of Time series and Causal models).
Please use the Self-service Tool to explore various Time-series Techniques. The Sample Data contains weekly Sales data for a Global Retail giant for 3 years.
Before using Time Series models, it is worthwhile to check presence of missing values, outliers etc in the data. Missing values etc. can be found by Search or Sort options in the Data View Tab. Also this tool allows the user to check the presence of outliers by qualitative means like Histograms, Line charts etc. Usually the Time Series is analyzed for presence of components like Level, Trend, Seasonal, Random etc. Once the statistical forecast has been generated using one of the statistical forecasting models the same can be reviewed by comparing the various KPIs for error. Once the statistical forecast is generated the consensus planning process fine-tunes forecast sales numbers. Consensus planning needs manual interventions.
Time Series refers to sequence of data, ordered over regular periods of time. Hence a Sales Data or Stock Price ordered in terms of weeks, months and so on can be called a Time Series Data. There are some specific techniques, which have been developed to analyse such data, and can be classified in following broad classes.
A raw Time-series data would show some stable explicable patterns like an increasing Trend or a Seasonal spikes. However it would have random components (Noise) too, which do not have any assignable factors for them. A prime motive is Time-series Analysis is to segregate these components, and untangle them from the noise.
Smoothing is one such technique, which averages out the wiggles or random components, so that more stable parts are more clearly identifiable. Various Moving Average techniques like Simple Moving Average, Exponential Moving Average etc. are used for this objective.
Simple Moving Average is an unweighted mean for the previous n data. n can take any value but greater than zero and less than or equal to the sample size. A Simple moving average technique is also a smoothing approach which uses equal weights across the time horizon.
However Exponential Smoothing techniques uses weights which exponentially decrease over a period of time. The most recent data has most weight but the data points further back in time are given less importance. Single, Double or Triple Exponential Smoothing Techniques are used. Single Exponential is used when there is only a Stationary or Level component evident in the time series, around which the data fluctuates. Double is used when in addition to Level, a Trend is also present. Triple exponential methods are used in forecasting when the time series data has trend and seasonality components too. Triple Exponential Exponential Smoothing is also called Holt-Winters Technique, and comes under a broad class of Time Series techniques called Decomposition Method.
Usually picking the best "order" is done by iterations and based on least error. However there are some automated techniques too, and are employed by Statistical Software to internally calculate them. the Mean Average Deviation(MAD) displayed as a bar chart in the Self-service Demand Forecasting tool may help in optimizing the order.
Decomposition is a techniques, which reveals the three components of a time series - Trend, Seasonality and Random(Noise). Specifying Frequency of Data is important in this Technique. Hence weekly data will have frequency of 52, quarterly data will have frequency of 4 and so on. In case of an error, increment or decrease the frequency little. The analysis helps in forecasting with appropriate options like without seasonal etc.
ARIMA is a very flexible and versatile technique, which can be applied on all kinds of Time-series data. It combines AR, MA and Differencing in single model. However its interpretation could be difficult and the accuracy depends on the choice of three parameters in it - Autoregressive, Differencing and Moving Average.