Forecasting Engine

Forecast quality is inevitably dependent on the quality of the underlying data. One-off events, such as a dramatic increase in sales or an unusual drop in demand, can dramatically change the forecast for the worse. Therefore, one must pay attention to the historic sales to produce accurate forecasts.

The forecasting module is designed to catch all possible trends, including slow and fast movers, seasonal trends, or items increasing or decreasing in sale. This means that users do not need advanced statistical knowledge as the system handles the forecasting part automatically. As sales data arrives, the forecasting module automatically calculates sales forecasts based upon one of the following five classes of forecasting models to satisfy all mainstream business forecasting requirements. The best fitting forecasting method gets chosen automatically for an item depending on the nature of the product and the amount of historic data available.

Historic sales or the time series may be classified based on ABC sales/volume analysis, this classification will be useful when reviewing and deciding whether judgmental adjustments should be made to the data to minimize the forecasting error. The ABC classifications are as follows;

A-items are usually 20% of the overall items that account for 80% of the sales, thus these items are of great importance to the firm. It is recommended to examine the forecasting error for these products on a regular basis and make judgmental adjustments to the sale history when appropriate.

B-items are approximately 30% of the overall items and account for 15% of the total sales. These items are therefore not as crucial to the firm as A items, the forecasting methods selected by the Forecasting Module are usually fairly accurate and human intervention is usually only required when exceptions occur in the sales.

C-items are of lowest volume and may include as many as 50% of the company’s items. Many of the C-items will have zero sales in the sales time series, with the occasional small sales and, more rarely, a large sale. The percent error of forecasting is likely to be high, which is why you would probably want to have a small confidence limit to keep your safety stock costs down. These items are inexpensive and only produce about 5% of the company’s sales, so for these items the cost of forecasting error is probably more acceptable than the cost of employees adjusting time series data. Thus, automation is key.

Expert Selection

Expert selection allows the Forecasting Module to select an appropriate univariate forecasting technique automatically. Expert selection operates as follows. If the data set is very short, Forecasting Module defaults to simple moving average. Otherwise Forecasting Module examines the data for the applicability of the intermittent or discrete forecast models. Although the forecasts produced from such models are just straight horizontal lines, they often provide forecasts superior to those from exponential smoothing for low- volume, messy data. If neither of these models are applicable to the data, the choice is now narrowed down to different forms of exponential smoothing and Box-Jenkins models. Forecasting Module next runs a series of tests on the data and applies a rule-based logic that may lead to a model selection based on data characteristics. If the rule-based logic does not lead to a definitive answer, Forecasting Module performs an out- of-sample test to choose between an exponential smohugo seothing model and a Box-Jenkins model.

Simple Methods

Simple Method includes moving average models and is for very short or extremely volatile data. This is a common inventory management method - used by wholesalers and distributors for demand forecasting - to average the sales over the previous few months. This method can work well for items in consistent demand, but it does not work so well for others. Because different items can have a very different demand pattern, it is extremely important to choose the most relevant forecasting method for each item.

If the data set is very short or has fewer than 10 points the Forecasting Module defaults to simple moving average.

Exponential Smoothing

Exponential smoothing works as its name suggests. It extracts the level, trend and seasonal indexes by constructing smoothed estimates of these features, weighting recent data more heavily. It adapts to changing structure but minimizes the effects of outliers and noise. Twelve different Holt-Winters Exponential Smoothing Models are provided to accommodate a wide range of data characteristics. Exponential smoothing models capture and forecast the level of the data along with different types of trends and seasonal patterns. The models are adaptive, and the forecasts give greater emphasis to the recent history verses the more distant past. The robustness of exponential smoothing makes it ideal when there are no leading indicators, and when the data is too short or volatile for Box-Jenkins.

The ‘wait-and-see’ attitude to changes around them is the intuitive way people employ exponential smoothing in their daily living.

Keep in mind that although exponential smoothing can take the following factors into consideration when projecting a forecast; the trend, level, seasonal effects, event effects, random events and noise. They do not and cannot include the effects of future random events or noise, so the forecast is much smoother than the actual future will turn out to be.

Discrete Distribution

These models apply to data consisting of small whole numbers, including some zeroes. The forecasts are non-trended and non-seasonal. Discrete distributions are for use on data that might consist entirely of zeroes and small integers. Infrequently used spare parts is an example of items that often fall into this class.

Although the forecasts produced are just straight horizontal lines, they often provide forecasts superior to those from exponential smoothing for low-volume, messy data.

Croston’s Intermittent Demand Model (Low Volume Model)

The Croston’s model is designed for data with numerous zeroes, like orders for a slow-moving part that is usually ordered to replenish stock. The non-zero data points are normally or log-normally distributed. The forecasts are non-trended and non-seasonal.

The time series consists of much sales data, especially for lower volume items with irregular demand. For many periods there is no demand at all. This might be the case for items that are usually ordered in batches to replenish downstream inventories. This method works by combining a smoothed estimate of the average demand for periods that have demand with a smoothed estimate of the average demand interval.

The forecasts produced will show straight horizontal lines.

Curve Fitting

Curve fitting identifies the general form of the curve that the data is following and is used to model the global trend of the historic sales data. Curve fitting is quite useful for short time series data, where the suggested minimum length is 10 data points. The curve fitting supports four types of curves: a straight line, quadratic, exponential, and growth (S-curve). Keep in mind that the curve does not accommodate seasonal patterns.

Box-Jenkins

Box-Jenkins works well for stable data sets and can capture and forecast both trend and seasonality. The data must consist of a minimum of 40 data points. The method is, quite simply, the richest family of statistical models that can be practically applied in the real world. Ideally, a forecaster would switch between Box-Jenkins and exponential smoothing models, depending on the properties of the data, which is precisely what the Forecasting Module automatic selection is designed to do. Box-Jenkins and Exponential Smoothing differ in that they are based on autocorrelations (stable data sets) rather than a structural view of level, trend and seasonality. Box-Jenkins models tend to perform better than exponential smoothing models for longer, more stable data sets and not as well for noisier, more volatile data.

Safety Stocks

Safety stock is kept to meet volatile demand during lead time (the time between placing an order and placing the units in stock), the Forecasting Module calculates the appropriate safety stock quantity based on the defined service level of a given product. Inventory control analysis requires the manager to balance the trade-off between inventory holding cost and the cost of a lost sale. Keep in mind that items with steady demand will have lower levels of safety stock than items with volatile demand.

The trade-off between holding cost and stock out cost can be seen in the following image.

Weekly and Daily Distribution

The Forecast Module uses monthly sales data to generate the initial forecast. The monthly forecast values are then distributed down to weekly and daily levels. The module assigns each week a specific sales amount based on the number of days the week has within the month. Then the weighted proportional sale for each weekday within each week is calculated, based on the last 30 weeks. This is for a general week, for a predefined number of weeks, the module calculates a special day distribution based on the historical distribution for that distinctive week such as (Easter week, week before Christmas, week before New Year’s etc.).

Keep in mind the following two rules:

1. If there are less than five data points, then the Forecasting Module switches to a simple moving average model.

2. If there is less than two years’ worth of data, then the Forecasting Module does not consider seasonal models.