Contributors Ruvisha Pillay and Bill Tonetti

In this article, we aim to give you an understanding of the statistical forecasting methods that are employed in Pivot Forecasting.

We will describe where to find and set methods, and where statistical outputs are displayed. We will also go into greater detail about the “Expert,” which is a rules-based AI solver that automatically selects the best methods and parameters. We will also discuss each of the other forecast types.

By default, every item in Pivot Forecasting will use the “Expert” method. When at the base level in your model, you can view the selected Forecast Type in the Visual Forecaster panel. You can change the selected method for each base-level item using the Visual Forecaster panel.

To change the selected method for multiple items at once, you must use the right-click Tree Menu and select “Update Forecast Parameters.”

The Forecast Type is the method that is requested. The Forecast Result is the method that is selected. Let’s explore that further.

In the example above, “Expert” is the requested Forecast Type. The Forecast Result is the “Level” exponential smoothing forecasting method which has been selected by the engine. The engine will not always do as requested if proper data is not available for that request. For example, if “Seasonal” is requested as the Forecast Type but the item doesn’t have enough historical data, the Expert will select a non-seasonal forecast type.

One of the predefined attributes of Pivot Forecasting is Forecast Results.

It is useful because it allows you to quickly use the Tree to view all of the items that utilize each forecasting method.

We’ve mentioned the Expert Forecast Type a few times, and you might be wondering how it actually works. Continue reading to find out!

First, the Expert analyzes each time series. Is there enough historical data to consider a seasonal method? If so, is there a consistent year over year pattern? Is demand concentrated over short selling seasons? Is the demand intermittent or consistent? Is it trending? The engine performs a number of tests to identify these and other characteristics.

Next, the engine will consider the user’s selected Forecast Type. The engine will attempt to use the requested method, but it may still have to bounce to a more appropriate method.

Once the engine has completed these basic classifications, it tries the most appropriate forecast types and compares the results. Within each forecast type, it runs an optimization to determine the parameters, solving for the best smoothing coefficients. Smoothing coefficients are what determines the added weight that is applied to the most recent levels, trends and seasonality.

Lastly, it compares the errors and other characteristics of each forecast type. The engine will not always choose the type that has the lowest error. This is because fit alone doesn’t determine the efficacy of a method for a given time series. Complex methods like ARIMA and seasonal exponential smoothing will almost always fit better but they may not predict future demand as well as simpler methods. Also at this stage, the engine uses an embedded Random Forecast (Machine Learning) classifier to assess the “riskiness” of a forecast. A risky forecast is one that seems to fit the data but probably will not predict well.

All of this may sound complicated - and it is - but that’s what it takes to make a world class forecasting engine!

To summarize, the Expert will:

Although there are thousands of possible algorithms, there are only 11 general Forecast Types in Pivot Forecasting. We take a closer look at each of them now.

The Expert will only use a seasonal method if there are at least two years of history. If the user selected the “Seasonal” Forecast Type, then the engine will consider seasonal forecasting methods with one cycle instead of two. It also relaxes the thresholds for the seasonal significance tests as well, but it may still revert to non-seasonal methods if necessary (i.e. if there is not enough history; if seasonality test thresholds are not met; if there is not enough recent history; or if the ML classifier deems it as being too risky.

To produce seasonal forecasts with highly random or incomplete history, we recommend that you use Seasonal Indexes instead of the Seasonal forecast type. In most cases, you can derive seasonality for an index using aggregated groups that demonstrate seasonality with greater confidence.

Both the Level and Level-Trend methods are in the family of exponential smoothing. They use either a 1-parameter or 2-parameter exponential smoothing method, with the second parameter being for trend. These are highly robust methods and they can be used on virtually any time series. As with the Expert, the engine will solve for optimal smoothing parameters. As previously mentioned, the smoothing parameters are the weights that are given to more recent activity.

Most trends in Pivot Forecasting are damped. The diagram above shows a highly damped trend, with the slope rapidly decreasing over time. Damped trends are more conservative than un-dampened ones. WIthout dampening, trends can result in many-fold increases to forecasted demand in later periods, which can be very risky for an automated forecast.

The non-seasonal forecast type will consider all non-seasonal methods including single and double exponential smoothing, moving averages and Croston’s. It will fall back to “None” if there isn’t any recent demand.

Croston’s is a smoothing technique that separately determines the probability that intermittent demand will occur, then multiplies that probably by the expected value of demand (removing the non-demand periods in the series). The result is a level forecast that is slower to respond and most suitable for low-velocity, intermittent demand.

The Moving Average model in Pivot Forecasting works on the average of 12 period, (or fewer if there are less than 12 periods). This is useful when a level forecast is needed that is less responsive to recent changes. It is also useful for intermittent series, particularly if there are negative values since the Croston’s forecast type cannot be used if there are any negative data points.

ARIMA stands for AutoRegressive Iterative Moving Average. ARIMA makes use of lagged moving averages to smooth time series data.

This forecast type can produce seasonal, trended, or non-seasonal forecasts. Sometimes, ARIMA will produce a “pulsed” forecast, which becomes more stationary over time. The top example shows an ARIMA forecast that’s both trended and seasonal. The bottom example shows a forecast that is “pulsed.”

The Seasonal Average method simply averages the previous two cycles. The Trended Seasonal Average method is the same, except it will also incorporate a damped trend. Both methods are relatively simple, calculating the forecast based on averaging “like periods” in multiple cycles. Examples of “like periods” would be months in a monthly model, or week numbers in a weekly forecasting model.

These forecast types are really useful for products like holiday-themed or highly seasonal foods, beverages, or garments.

The “None” forecast type is actually used often, because it's the forecast type of choice when an item does not need to be forecasted. This forecast type is most useful for end of life products or products that you only wish to sell as buy-to-order.

In summary, there are a variety of forecasting methods in Pivot Forecasting and a world class Expert and user interface for supporting the full range of demand forecasting requirements. And, we do it all to help you make better plans with Netstock.

Have you read the related articles? Check out our Pivot Forecasting collection!