Module 2 - Section B - Forecasting Techniques

Absolute Percentage Error(APE)

expressing forecast error as a percentage

Curve fitting

An approach to forecasting based on a straight line, polynomial, or other curve that describes some historical time series data.

Additive Seasonal Variation

Additive seasonal variation assumes that there is a constant seasonal amount regardless of what the trend or average amount is. The seasonality is overlaid on the trend.

Qualitative Forecasting Techniques

An approach to forecasting that is based on intuitive or judgmental evaluation. It is used generally when data is scarce, not available, or no longer relevant.

Focus Forecasting System

A system that allows the user to simulate the effectiveness of numerous forecasting techniques, enabling selection of the most effective one.

Tracking Signal

Used to alert forecasters that a variation has exceeded a preset limit or multiple of inherent randomness as expressed in the MAD is the tracking signal, which is the sum of the forecast deviations (not absolute) divided by the MAD. it is compared to a threshold or trip value that can provide early warning for when bias or other types of forecast errors are becoming problematic.

within allowable item tolerances (WAIT)

WAIT is derived from what makes an incorrect forecast correct. Organizations need to determine how much error creates a problem. If it is 10 percent, then that is the delta. Each item is allowed a tolerance of plus or minus 10 percent. If it is within the tolerance, it is called a HIT; if it is outside the tolerance, it is a MISS.

Multiplicative Seasonal Variation

With multiplicative seasonal variation, the trend is multiplied by the seasonal factors. The highs and lows would both increase in this method, and the seasonality would appear to become more volatile.

Three types of causal techniques

simple regression, multiple regression, and econometrics.

Econometric Model (Multiple Regression Models)

A set of equations intended to be used simultaneously to capture the way in which dependent and independent variables are interrelated.

Econometric Modeling

A set of equations intended to be used simultaneously to capture the way in which dependent and independent variables are interrelated.

First Order smoothing or single exponential smoothing

A single exponential smoothing; a weighted moving average approach that is applied to forecasting problems where the data does not exhibit significant trend or seasonal patterns.

Collaborative Planning, Forecasting, And Replenishment (CPFR)

"A collaboration process whereby supply chain trading partners can jointly plan key supply chain activities from production and delivery of raw materials to production and delivery of final products to end customers."

Outlier

"A data point that differs significantly from other data for a similar phenomenon. For example, if the average sales for a product were 10 units per month, and one month the product had sales of 500 units, this sales point might be considered an outlier."

Time Series Forecasting

"A forecasting method that projects historical data patterns into the future. [It] involves the assumption that the near-term future will be like the recent past". The methods used in time series forecasting are less complex mathematically and thus easier to explain to decision makers. Time series methods assume that the factors that influenced the past will continue on into the future. When that trend is unlikely to be stable, causal/associative forecasting may be needed.

Adaptive Smoothing

"A form of exponential smoothing in which the smoothing constant is automatically adjusted as a function of forecast error measurement."

Standard Deviation

"A measurement of dispersion of data or of a variable. The standard deviation is computed by finding the differences between the average and actual observations, squaring each difference, adding the squared differences, dividing by n - 1 (for a sample), and taking the square root of the result."

Decomposition

"A method of forecasting where time series data is separated into up to three components—trend, seasonal, and cyclical—where trend includes the general horizontal upward or downward movement over time; seasonal includes a recurring demand pattern such as day of the week, weekly, monthly, or quarterly; and cyclical includes any repeating, nonseasonal pattern. A fourth component is random—that is, data with no pattern. The new forecast is made by projecting the patterns individually determined and then combining them."

Seasonal Index

"A number used to adjust data to seasonal demand."

Normal distribution

"A particular statistical distribution where most of the observations fall fairly close to one mean, and a deviation from the mean is as likely to be plus as it is to be minus. When graphed, the normal distribution takes the form of a bell-shaped curve."

Sample

"A portion of a universe of data chosen to estimate some characteristics about the whole universe. The universe of data could consist of sizes of customer orders, number of units of inventory, number of lines on a purchase order, and so forth."

Seasonality

"A predictable repetitive pattern of demand measured within a year where demand grows and declines. These are calendar-related patterns that can appear annually, quarterly, monthly, weekly, daily and/or hourly."

Delphi Method

"A qualitative forecasting technique where the opinions of experts are combined in a series of iterations. The results of each iteration are used to develop the next, so that convergence of the experts' opinions is obtained." The Delphi method has had good success at arriving at reliable forecasts, but it is time-consuming and labor-intensive. It is often used only for strategic-level estimation.

Life Cycle Analysis (LCA)

"A quantitative forecasting technique based on applying past patterns of demand data covering introduction, growth, maturity, saturation, and decline of similar products to a new product family." This is a comparative technique that analyzes and adapts existing data and patterns and extrapolates them to create a forecast based on historical information.

Demand Filter

"A standard set to monitor sales data for individual items in forecasting models. Usually set to be tripped when the demand for a period differs from the forecast by more than some number of mean absolute deviations."

Base Series

"A standard succession of values of demand-over-time data used in forecasting seasonal items. This series of factors is usually based on the relative level of demand during the corresponding period of previous years. The average value of the base series over a seasonal cycle is 1.0. A figure higher than 1.0 indicates that demand for that period is higher than average; a figure less than 1.0 indicates less-than-average demand. For forecasting purposes, the base series is superimposed upon the average demand and trend in demand for the item in question."

Probability distribution

"A table of numbers or a mathematical expression that indicates the frequency with which each of all possible results of an experiment should occur."

Exponential Smoothing Forecast

"A type of weighted moving average forecasting technique in which past observations are geometrically discounted according to their age. The heaviest weight is assigned to the most recent data. The smoothing is termed exponential because data points are weighted in accordance with an exponential function of their age. The technique makes use of a smoothing constant to apply to the difference between the most recent forecast and the critical sales data, thus avoiding the necessity of carrying historical sales data. The approach can be used for data that exhibits no trend or seasonal patterns. Higher order exponential smoothing models can be used for data with either (or both) trend and seasonality."

Moving Average

"An arithmetic average of a certain number (n) of the most recent observations. As each new observation is added, the oldest observation is dropped. The value of n (the number of periods to use for the average) reflects responsiveness versus stability in the same way that the choice of smoothing constant does in exponential smoothing. There are two types of moving average: simple and weighted."

Weighted Moving Average

"An averaging technique in which the data to be averaged is not uniformly weighted but is given values according to its importance."

Time Series Analysis

"Analysis of any variable classified by time in which the values of the variable are functions of the time periods. Time series analysis is used in forecasting. A time series consists of seasonal, cyclical, trend, and random components."

Extrapolation

"Estimation of the future value of some data series based on past observations. Statistical forecasting is a common example."

Smoothing Constant

"In exponential smoothing, the weighting factor that is applied to the most recent demand, observation, or error. In this case, the error is defined as the difference between actual demand and the forecast for the most recent period. The weighting factor is represented by the symbol α. Theoretically, the range of α is 0.0 to 1."

Mean

"The arithmetic average of a group of values."

Sampling distribution

"The distribution of values of a statistic calculated from samples of a given size."

Median

"The middle value in a set of measured values when the items are arranged in order of magnitude. If there is no single middle value, the median is the mean of the two middle values."

Mode

"The most common or frequent value in a group of values."

Forecast Management

"The process of making, checking, correcting, and using forecasts. It also includes determination of the forecast horizon."

Correlation (Causal Techniques)

"The relationship between two sets of data such that when one changes, the other is likely to make a corresponding change. If the changes are in the same direction, there is positive correlation. When changes tend to occur in opposite directions, there is negative correlation. When there is little correspondence or changes are random, there is no correlation."

Distribution Of Forecast Errors

"The tabulation of the forecast errors according to the frequency of occurrence of each error value. The errors in forecasting are, in many cases, normally distributed even when the observed data does not come from a normal distribution."

Regression Analysis

A statistical technique for determining the best mathematical expression describing the functional relationship between one response and one or more independent variables.

Benefit of Mean Absolute Percent Error (MAPE)

A drawback of the MAD and MSE calculations is that they are not meaningful unless you know what an expected result should be based on the size of the forecast. The mean absolute percent error (MAPE) is useful as it shows the ratio, or percentage, of the absolute errors to the actual demand for a given number of periods. If MAPE is greater than a certain percentage, additional review is required.

Pyramid forecasting

A forecasting technique that enables management to review and adjust forecasts made at an aggregate level and to keep lower-level forecasts in balance.

Multiple Regression Models

A form of regression analysis where the model involves more than one independent variable.

When to use Exponential Smoothing

A method is often used because of its flexibility and low cost. Organizations use this method when they want to minimize the lag that exists when trends shift, but, like all time series models, it cannot eliminate this lag.

Second Order smoothing or double smoothing

A method of exponential smoothing for trend situations that employs two previously computed averages, the singly and doubly smoothed values, to extrapolate into the future.

Bias

Bias is a consistent deviation from the mean in one direction (consistently high or consistently low).

How to correct Bias ?

By better aligning forecasts with demand, better anticipating new customer orders and product demands, and developing better estimates of safety stock (and, as a result, reducing inventory investment).

Two primary variables dictate the best option:

Complete versus incomplete data. If all the required sales data for a particular item as well as the causal variables are available, then the data are considered complete. Incomplete data would have limited—or even an absence of—sales data for a particular product or not have the causal variables identified. Stable versus unstable data. Stable data have a distinct pattern such as seasonality or trends. There's a randomness and no distinct pattern to unstable data.

Difference between correlation and causation

Correlation is an observation that the change in an independent variable has a measurable effect on a dependent variable. However, just because the effect can be reliably observed over time does not mean that the one thing caused the other thing. It could be that some third force is affecting both of them, or the correlation could be a coincidence. It is important that forecasters apply common sense when selecting independent variables (predictors), or decision makers might reject the forecast results.

Least-Squares Method (Simple Regression) (Linear Regression)

Method of curve fitting that selects a line of best fit through a plot of data to minimize the sum of squares of the deviations of the given points from the line.

In Delphi Method, Anonymity is used for two reasons

First, it helps prevent dominant personalities from influencing the group opinion. When this "groupthink" effect is in play, otherwise independent thinkers might become emotionally committed to an unrealistic forecast. The other problem anonymity prevents is a "stake in the ground" mentality—when a person has already publicly committed to a forecast result and doesn't want to lose face by changing his or her declared position. Since the Delphi method is anonymous, it is easier to change a position given more information.

Quantitative forecasting techniques

Include causal, time series (including moving averages, exponential smoothing, and decomposition), and life cycle analysis. Pyramid forecasting, as just discussed, may rely partly on qualitative and partly on quantitative methods.

Benefit of Mean Square Error (MSE)

Note that the process of squaring each error gives you a much wider range of numbers. (3.4 squared is 11.56, while a 0.63 error rate squared would become 0.397.) The greater range gives you a more sensitive measure of the error rate, which is useful if the absolute error numbers are relatively close together and the reduction of errors is important. Measuring the extent of deviation helps determine the need to improve forecasting or rely on safety stock to meet customer service objectives.

Standard Deviation Benefits

Standard deviation measures the difference between period actual demand and average demand (specifically, the mean demand) during a forecast horizon (for example, 10 weeks). It is not a measure of forecast error and does not use forecast error as part of the calculations. Rather, it is a measure of how far results tend to deviate from the average result. Because it indicates demand variability, assuming that standard deviation is distributed normally (i.e., on a bell curve), it can be used to develop insights into calculating safety stock and improving customer service and inventory management.

The variables used in causal techniques can come from internal or external data sources.

The predictor is called the independent variable; the element being predicted is called the dependent variable. These techniques are best for long-term forecasting at the aggregate level.

Mean Absolute Deviation (MAD)

The average of the absolute values of the deviations of observed values from some expected value. [This] can be calculated based on observations and the arithmetic mean of those observations. An alternative is to calculate absolute deviations of actual sales data minus forecast data. This data can be averaged in the usual arithmetic way or with exponential smoothing.

How do we tell whether there is enough correlation to use this predictor in our sales forecast?

The coefficient of correlation (r) is a statistical term for the strength of correlation, and in this case it is +0.793, which is a strong positive correlation worth pursuing. That is, this predictor explains about 79 percent of the change in roofing sales. Weak r values would indicate that you should select a different predictor.

Forecast Accuracy

The complement of the forecast error as a percentage

When to use simple moving average

The simple moving average can be useful when demand is relatively constant from period to period. The method can be used to prevent an overreaction to a random or irregular spike or dip in a given month because it smooths out these variations. However, if there is a change in a trend, this method would be slow to respond to it. It would lag the trend, in other words.

Time Series Forecasting Methods

There are a number of types of time series forecasting, ranging from the very simple to the relatively complex. Naive forecasting simply assumes that the last period's demand will be this period's forecast. It can be cost-effective but does not account for trends, and any random spike or trough in demand would be carried forward. Moving averages, exponential smoothing, and decomposition are also time series methods.