Business Forecasting: An Annotated Bibliography

Introduction

 

The field of forecasting includes supply forecasting, such as for agricultural commodities or the oil industry, and extends to economic forecasting, such as GDP forecasts produced by the Federal Reserve Bank.  The six articles cited in this annotated bibliography are published in The Journal of Business Forecasting (JBF), a publication of the Institute of Business Forecasting and Planning.  Established in 1982, JBF is a peer-reviewed academic journal edited by Dr. Chaman Jain, Professor of Economics at St. John’s University.  Articles in the journal are written for corporate demand planners and S&OP professionals. The field of business demand forecasting is dynamic, and practitioners need to stay abreast of current trends in the field.  In his letter to readers in the current issue of the journal (Winter 2017-2018), Dr. Jain comments on the expanding role of e-commerce, the advent of automated forecasting software, and the growing need for demand planners to adopt the increasing availability of sophisticated analytical tools.

Forecasting methods are covered in chapter four of Heizer and Render’s (2014) textbook, Operations Management: Sustainability and Supply Chain Management (104-136).  The articles annotated here build on that material and seek to create general familiarity with the field of business forecasting.  Using an evaluative annotation approach, each article has been given an extended exposition of its content. The accompanying PowerPoint presentation, rather than introducing articles alphabetically by author (mea culpa to APA), approaches the bibliography in a thematic sequence. Beginning with an overview of forecasting models, the topics cover the use of external data and economic indicators, matching forecasting methods to product segments, the impact accurate forecasts have on inventory and financial metrics, and talent management as an important part of the forecasting process.  The annotated bibliography that follows is alphabetical by author’s last name.

 

Chase, Charles W. (Winter 2015-2016). The Importance of Product Segmentation. The Journal of Business Forecasting; 34, 4, 36-40.

Charles Chase is an expert in sales forecasting, market response modeling and supply chain management; he has worked for companies such as Johnson & Johnson, Coca-Cola, and Heineken, among others (p. 36). Companies are understanding that not all their products are forecastable, says Chase, and “are asking themselves what is forecastable and what is not, and how can they segment their products to get the most accuracy across their product portfolio” (p. 37). Forecastability is impacted by consumer demand and data constraints. Demand planners must segment their brands and products and apply the correct forecasting method to obtain the best performance across their product lines.

Many practitioners realize poor forecast performance simply because they use one methodology to forecast all of their products (p. 37). The author states that in many cases the forecast accuracy is less than 70%, whereas desired results would be in the 85% to 95% range (p. 37). Product groups have different data patterns, depending on their life cycles and marketing and sales support. When demand planners “cleanse” a product’s historical demand data, they also lose data on sales promotion responses and seasonality, which can result in their ERP (Enterprise Resource Planning) system picking a moving average model instead of a seasonal exponential model (p. 37). The right model can improve forecasts by 10% across the firm’s product portfolio.  Chase recommends decomposing each data set by product or brand to find the trend, seasonality, cyclical variance, and unexplained error (p. 37). This kind of time series analysis can uncover true demand patterns.

The value that Chase brings to demand planning and forecasting is his concept of segmenting products and brands into four quadrants and then overlaying the best forecasting method over each quadrant to achieve better forecast results.  The four general product quadrants are: 1) low value, low forecastability, 2) low value, high forecastability, 3) high value, low forecastability and 4) high value, high forecastability (p.37). He then provides more specific product characteristics for each quadrant: 1) slow moving, 2) new products, 3) fast moving, and 4) steady state (p.37). The table below, reproduced from the article, shows these four quadrants and product portfolio management principles:

New Products                 High Value                         Low Forecastability Product Line Extensions New Products                            Short Life Cycle Products High Priority Products              Strong Trend                  Seasonal Fluctuations             Possible Cycles                 Advertising Driven                 Sales Promotions Fast Moving                        High Value                          High Forecastability
Company Value    
Slow Moving Products                         Low Value                                           Low Forecastability Low Priority                             (Regional Specialty Products)                              Some Trend                             Seasonal Fluctuations Intermittent Data Low Priority Products               Strong Trend                                Highly Seasonal                          Possibly Cycles                           Minor Sales Promotions Steady State Products             Low Value                        High Forecastability
Low    <——————      Forecastability  —————>  High

 

The appropriate forecasting methodologies can also be segmented into four quadrants and matched with the segmented product characteristic quadrants shown above. For example, slow moving, low forecastability products can only be approximately forecasted using some trend and seasonal fluctuation analysis and moving averaging, whereas high value, high forecastability products show strong trend and season fluctuations and are responsive to sales promotions and advertising (p. 38). This group of products can be further analyzed using causal models (dynamic regression, multiple linear regression) to reduce unexplained variance.  Chase’s breakout of statistical methods into four corresponding quadrants (not shown here) can then be overlaid on top of the product characteristics segmentation table, above. In brief, new product demand can be forecasted using judgmental methods, clustering and data mining. Slow moving products can benefit from moving averages, sales force composites and intermittent demand models. Causal methods work well for the fast-moving products quadrant, and time series methods such as exponential smoothing work well for the steady-state products quadrant (p. 39).

Chase underscores the point that more sophisticated statistical methods are needed to improve on time series methods which are useful for segmenting product group demand history into basic trend, seasonality, cyclical and unexplained variance (p. 40). Heizer and Render (2014) explain time series methods in chapter four of Operations Management (pp. 108-125).  Chase goes the additional step in this article by mapping the most applicable statistical and forecasting methods against the segmented product characteristics framework to help demand planners achieve better forecasts.

 

Chase, Charles W. (Spring 2016) Forecast Accuracy Has No Impact on Inventory! Really? The Journal of Business Forecasting, 35, 1, 26-29.

Charles Chase is an expert in sales forecasting, market response modeling, and supply chain management. Currently with the SAS Institute, Inc., he has worked for companies such as Johnson & Johnson, Coca-Cola, and Heineken (p. 26).  Chase’s main argument here is that companies have sorted themselves out into two supply chain camps: demand or supply. In his opinion, firms have swung back and forth between these two camps; in recent years, companies have abandoned supply chain strategies because the use of buffer inventory is not as cost efficient as it was (p. 26).  He states, “although poor forecasting has been identified as the root cause, companies continue to use traditional statistical methods like moving averaging and non-seasonal exponential smoothing models, which are only accurate one period into the future” (p. 26).

There is a lot of room for forecasting accuracy. Based on his work with 100 companies over 10 years, “the average forecast accuracy is between 50-65 percent at the aggregate level, and between 35-45 percent at the lower mix levels” (p. 27).  A classic cause of poor demand forecasts is the use of moving average and non-seasonal exponential smoothing models that are only accurate a few periods into the future. Upper and lower forecast ranges tend to be cone shaped beyond that point, whereas more advanced statistical methods like ARIMA and dynamic regression are more accurate farther out (p. 28).  The use of better models translates into lower safety stock. Often businesses’ ERP (Enterprise Resource Planning) software only supports the simpler methods. The author says that demand forecasting and planning has received “little attention and investment in people, analytics, and technology over the past decade” (p. 28).

Companies have shown 10-30 percent improvements in forecast accuracy by using holistic modeling based on predictive analytics. Businesses often “cleanse” their products’ demand history by breaking it down into two streams: 1) baseline, and 2) promoted.  The promoted stream is actually a combination of seasonality and promoted volume, which can be spiky.  The baseline stream tends to be a moving average. Demand planners often try to put these two data streams back together with “the result, 1+1 now equals 5” (p. 29). Advanced analytics for supply planning together with better forecasts creates a synergy effect of an additional 15-30 percent reduction in finished goods inventory (p. 29). The result is reduced inventory costs, increased revenues and profit, and more available working capital (p. 29).

Companies that are moving toward becoming more demand-driven are doing better with demand forecasts and reducing inventory safety stock. New forecasting models, sometimes referred to as “Consumption Based Modeling”, link downstream data to upstream data using a process called “Multi-Tier Causal Analysis” (MTCA) (p. 29). Heizer and Render (2014) clarify “downstream” as distributors and retailers, and “upstream” as suppliers (p.447). Chase references his own Spring 2015 article in The Journal of Business Forecasting for more information about the correlation between upstream and downstream data. By not relying entirely on either demand or supply forecasts businesses can be better at solving their supply chain challenges. A holistic view of supply and demand can lead to more success in building a responsive supply chain. Heizer and Render (2014) cover forecasting methods in chapter four of Operations Management and explain supply chain logistics and inventory models in chapters 11 and 12, respectively.

 

Homareau, Jack. (Fall 2015) Forecasting Sales Volumes with Economic Indicators. The Journal of Business Forecasting, 34, 3, 32-34.

The author brings many years of macroeconomic modeling and forecasting experience to this article, which outlines a methodology for forecasting sales volume using economic indicators.  Specifically, he showed how to create a baseline product forecast for US automobile sales using U.S. housing starts data as the economic driver.  The Housing Starts data series correlates well with automobile sales volume, and along with product promotions and marketing plans can improve sales forecasts. He also evaluated the forecasting model’s performance. The sales data used in this example are quarterly U.S. auto sales from 1999 through 2012. Actual sales for 2013 were used to evaluate the model’s predictions for 2013 sales. The methodology used was a simple regression model, with housing starts as the explanatory variable and auto sales as the response variable (pp. 31-32).   For 2013, actual U.S. housing starts (rather than projected U.S. housing starts) were used to forecast auto sales over the same period. Data from the U.S. Census Bureau was converted from monthly to quarterly totals.

Homareau’s next step was to calculate the quarterly percentage change in both data series on a year-over-year (YoY) basis. Then he developed the regression model, where x is the YoY quarterly percentage change in housing starts, and y is the YoY quarterly percentage change in U.S. auto sales (pp. 33-34). The resulting coefficient of correlation is 0.74, and the regression line equation is Y=1.1 + 0.5 x. The third step was to prepare 2013 forecasts. The forecasted auto sales growth rate for 2013 Q1 was calculated at 18.1%, based on U.S housing starts growth of 33.9% in Q1 of 2013.  Automobile sales growth rates for Q1- Q4 of 2013 are calculated below (p. 34).

Period Housing Starts Growth Rate Vehicle Sales Growth Rates (YoY: quarterly)
2013 Q1 33.9 1.1 + (.05 x 33.9)= 18.1
2013 Q2 17.0 1.1 + (.05 x 17.0) = 9.6
2013 Q3 13.0 1.1 + (.05 x 13.0) = 7.6
2013 Q4 12.9 1.1 + (.05 x 12.9) = 7.6

 

This growth rate is multiplied against 2012 Q1 actual sales to obtain projected 2013 Q1 sales of 4,179.9. The following table shows these calculations and 2013 sales projections (p. 34).

Period Motor Vehicles Sales (000) Projected YoY: Quarterly Growth Rates Projected Quarterly U.S. Motor Vehicles Sales (000)
2013 Q1 3,539.3 (2012-Q1) 18.1 3,539.3 x 1.181  = 4,179.9
2013 Q2 3,883.1 (2012-Q2) 9.6 3,883.1 x 1.096  =  4,255.9
2013 Q3 3,699.5 (2012-Q3) 7.6 3,699.5 x 1.076  = 3,980.7
2013 Q4 3,665.8 (2012-Q4) 7.6 3,665.8 x 1.076  = 3,944.4

The final step in the process was to compare forecasted sales with actual sales, and compute the MAPE (Mean Absolute Percentage Error). For 2013 Q1, actual sales were 3,754.6 and projected sales were 4,179.9. The forecast error here is 425.3, with a MAPE of 10.2%. The following table shows these results as well as the calculated overall 2013 MAPE of 2.9% (p. 34).  The author

Period Projected Sales (000) Actual Sales (000) Absolute Forecast Error Absolute % Error
2013 Q1           4,179.9                 3,754.6 425.3 10.2
2013 Q2           4,255.9                 4,210.3 45.6 1.1
2013 Q3           3,980.7                 4,028.8 48.1 1.2
2013 Q4           3,944.4                 3,890.0 54.4 1.4
MAPE       2.9

 

states this error rate is fairly good by industry standards, but that an additional economic variable and some manual adjustments by a person with “domain knowledge” (auto industry sales data background), would have improved the projected sales for 2013 and reduced forecast error.

Heizer and Render (2014) cover regression analysis and MAPE in chapter four of Operations Management (pp. 106-132). It was helpful to see Homareau use quarterly percentage change in the data series in his regression rather than actual auto unit sales regressed on the actual number of housing starts. The use of percentage growth rate will be a helpful data transformation tool in other cases where the explanatory and dependent variables may be measured on different scales.  I entered all of the author’s data into Excel and ran the regression. I can confirm the coefficient of correlation is 0.74; however, the regression model is more accurately stated as Y = 1.1 + 0.455 x (the author rounds it up to Y= 1.1 + 0.5 x). The more accurate coefficient will result in a smaller forecast error. See the Excel “Summary Output” below for more results. This article will be quite helpful for students who are learning how to forecast business data using a simple regression model that includes U.S. economic data as the explanatory variable (independent variable).

SUMMARY OUTPUT
Regression Statistics
Multiple R 0.736882537
R Square 0.542995873
Adjusted R Square 0.53385579
Standard Error 8.227878293
Observations 52
ANOVA
  df SS MS F Significance F
Regression 1 4021.815363 4021.81536 59.4082023 4.72186E-10
Residual 50 3384.89906 67.6979812
Total 51 7406.714423      
  Coefficients Standard Error t Stat P-value Lower 95% Upper 95%
Intercept 1.115357672 1.158543528 0.96272401 0.34031911 -1.211645488 3.442360833
Housing YoY % Change 0.455566491 0.059105591 7.70767165 4.72186E-10 0.336849417 0.574283565

 

 

Jain, Chaman. (Winter 2006/2007) Benchmarking Forecasting Models. The Journal of Business Forecasting, 25, 4, 14-17.

The purpose of this article is to help business forecasters determine how to apply the right model to their data sets, taking into consideration that data sets have patterns. The author describes three types of forecasting models, identifies fundamentals of modeling, discusses the types of models used in business, and concludes with model selection.  The three types of models are: 1) Time Series, 2) Cause-and Effect, and 3) Judgmental (p. 14). Time Series models are based on the assumption that the future will resemble the past. They work best for short-term forecasting due to data stability in the short run. They include: 1) Simple- and Moving Averages, 2) Trend, 3) Exponential Smoothing, 4) Decomposition, and 5) Autoregressive Integrated Moving Average (ARIMA) (p. 14).  Cause-and- Effect Models involve independent and dependent variables, where the average relationship between the two can be projected into the future. Jain provides examples: 1) Regression, 2) Econometric, and 3) Neural Network. Judgmental Models can be used when there is no historical data, when the market is volatile, and/or when a very long-term forecast is contemplated (p. 14).

Jain offers nine fundamentals of models: 1) Datasets form patterns and the model should attempt to capture the pattern as best as possible. The actual model will include pattern plus error. 2) 100% accuracy is not necessary.  Minimization of error is the goal. 3) More data are not necessarily better. In the consumer products industry, most companies use data of 30 months or less. 4) Sophisticated models are not always better. Start with a simple model and move to more complex models. 5) There is no magic model. Change in the real world means models need to adapt to the datasets being analyzed. 6) Models age with time. Over time, data is usually dynamic and models will not always work as they used to. 7) Each model has its own data requirements. Some models require 35 or more observations to prepare a forecast. 8) Statistical forecasts are nothing more than baseline forecasts. The art and science of forecasting may require a judgmental overlay. 9) Forecasts should not be prepared in isolation. It is important to consult stakeholders such as marketing or finance people (pp. 15-16).

Based on a 2006 survey, Time Series models are used by 72% of industries, Cause-and Effect by 17% and Judgmental by 11% (p. 16). Jain reports that within Time Series model usage, Averages and Simple Trend are used 60% of the time, Exponential Smoothing 30%, and other methods are used 10%. Cause-and-Effect models are being used more often in business forecasting, with Regression used the most. Judgmental models that are used most are Survey (50%), followed by Analog at 27% (p. 17). Jain provides bar charts that show all the model usage survey results.  Software systems often have an internal “expert system” that select the best model for the application. An example of a criterion might be minimizing Mean Absolute Percentage Error (MAPE). Forecasting software systems often are based on time series models and some provide an expert system in regression. Forecasters can use suggested software tools or select their own model preference. Jain recommends starting by choosing an acceptable error rate and then using a simple model before progressing to complex models (p. 17). Monitoring the model’s performance is important, as is using “ex post forecasts” to determine if the model is good.

Heizer and Render (2014) present this material in the forecasting chapter of Operations Management (pp. 104-136). While Jain’s article is more than 10 years old now, it is still a valuable introduction for students who are new to forecasting. Knowing how often industries use time series models (specifically moving averages and exponential smoothing), or regression models, is helpful.  Jain’s discussion of modeling fundamentals is good for putting models in better perspective, and for providing some guidance regarding model selection based on dataset patterns. The article could elaborate more on available forecasting software options, although even the brief discussion here is helpful for understanding some of the software capabilities and tools that vendors provide.

 

Wagner, Rich. (Spring 2016) Why External Data Are Vital to Demand Forecasts.  The Journal of Business Forecasting, 35, 1, 30-34.

Rich Wagner is president and CEO at Prevedere, a firm that helps companies improve their forecasts. Most companies use only internal historical data and could benefit from incorporating external factors in their predictive analytics.  External influences have a large impact on firm performance but many businesses rely exclusively on internal historical financial data. Wagner recommends that organizations assess external data sets to gauge opportunities or avoid risky outcomes. Algorithms are now included in predictive modeling software to help integrate external data sets with internal data. Unemployment data, hourly earnings, and consumer sentiment are examples of external data that can guide decisions about product demand.

Companies can improve their strategic course and develop better business plans by paying attention to their external world. To use Wagner’s example, knowing that crude oil prices are in general decline, that China’s growth is slowing, or that the European Union is facing economic instability can help firms protect shareholder value by using this intelligence ahead of time (p. 31). Changing consumer spending and buying habits can cause companies to close stores and lose revenue if these trends are not responded to in time. Information such as disposable income levels, U.S. housing starts and regional trends can help companies take advantage of opportunities they may otherwise miss.

According to an analysis of corporate forecasts by The Economist Intelligence Unit, only one percent of annual forecasts are accurate, and quarterly forecasts are accurate 13 percent of the time (p. 31). Failing to account for macroeconomic factors has cost public companies $200 billion a year (KPMG study) (p. 32). Relying on internal data and subjective judgments can be costly. Predictive modeling tools use algorithms to match specific relevant data to an organization’s business goals. Wagner gives an example of a U.S. operator of more than 650 convenience stores and gasoline stations that adjusted their purchasing decisions and made better decisions about inventory after incorporating leading economic indicators and weather data into their demand forecasts. Another example is that of a large beverage company improving demand forecasts for its China market by 20 percent, resulting in reduced inventory costs. Each five percent error in demand required an extra two percent in safety stock for this company (p. 33).

The article outlines a six-step process to generate forecasts:  1) Collect the right information. Systematically merge financial data with leading economic and geopolitical drivers. 2) Marry internal and external data to identify key drivers. Go beyond the Conference Board’s Composite Index of Leading Indicators, which is usually two months old. 3) Implement predictive data analytics. Apply advanced pattern-matching and technology solutions to get insights. 4) Use the information to predict the future. Companies that place value in leading economic indicators achieve a 5.14 percent higher return on equity, according to a Garner and Wharton School study. 5) Put the information in the right hands. Share the information across corporate functional and business line users. 6) Make it repeatable. Companies need to update their forecasts regularly and find their leading drivers instead of relying on one-off economic consulting services (pp. 33-34).

Although the article at times may come across as being a promotional piece for the author’s business, it does emphasize the critical nature of not creating business forecasts in isolation from external data such as geopolitical and economic data. By meshing external data with their internal metrics such as EPS (earnings per share), ROE (return on equity), ROI (return on investments), EVA (economic value added) and EBITDA (earnings before interest, taxes, depreciation, and amortization), companies can be considerably more accurate in their business forecasts (p. 30). More refined forecasts are more likely to predict the future more accurately. Better accuracy can mean reduced inventory costs, improved sales, higher profit margins, and increase shareholder value and stock prices.

 

Wilson, Eric and Breault, Jason. (Spring 2016). Improving Forecast Accuracy through Talent Management.  The Journal of Business Forecasting, 35, 1, 4-9.

Eric Wilson is a Certified Professional Forecaster (CPF) and a member of the Institute of Business Forecasting’s (IBF) Board of Advisors. Jason Breault is the managing director of a management recruiting firm that identifies talent in demand and supply planning, and sales and operations planning (S&OP). Their article looks at the people side of demand planning and the four key pillars of better talent management. They argue that practitioners think about process, and software companies talk about technology, but more attention needs to be paid to people – an important part of the forecasting process (p. 5).

Supply chains have changed to be driven more by demand, and the demand planning role needs to change also. Businesses that do not improve their demand planning talent are at risk of falling behind in their capabilities. A talent management strategy is required: 1) talent based on competency, 2) career-path culture and visibility inside the firm, 3) training and development, and 4) performance-based management (p. 5). Based on a Supply Chain Insights survey, Demand Planner, Supply Planner, Manager of S&OP, and Director of Supply Chain Planning are hard positions to fill (p. 5). One reason is that often people that are analytical and forward-thinking do not have the soft skills needed to relate to sales and marketing people. They suggest that these roles can be broken down into: 1) demand analyst responsible for statistical baseline forecasts, and 2) demand planner who can interact with sales and marketing and “instill a sense of partnership with them” (p. 6) Other roles they mention in Product Lifecycle Management (PLM), is Launch Leader, and Product Innovation Planner.

Communicating to the broader organization is important in creating visibility for these roles. Knowing how positions interrelate is important. Key stakeholders in organizations need to be sold on the role of the demand planning function within their supply chain process (p. 6). “In the land of the blind, the one-eyed man is king,” they say, illustrates the need to forecast and plan forward (p. 6). People in these roles also must work to build their department’s brand within the organization at large. Companies can also require certifications such as from the IBF (Institute of Business Forecasting), to ensure their forecasters have the necessary knowledge and skills. Certification in supply chain can help shorten a company’s acquisition of the knowledge base needed to improve results. Survey have shown that those with professional certification are more likely to encourage consensus and build critical relationships to improve performance (p. 7).

Demand planners within a firm need to expand the yardstick against which they are measured. One measurement method is FVA% (Forecast Value Add). This concept will help determine which steps in the forecasting process are improving forecasting accuracy. Statistical baseline forecasts can be measured against a naïve forecast, or planners’ overrides of the baseline forecast, thereby identifying inputs and processes that are effective (p. 9). The company will have a better feel for what drives their forecasts’ accuracy and achievement of project milestones. Industry leaders have focused on talent development, creating a culture for success, employee competency, and measurement of the right performance indicators (p .9). Managers in these companies not only understand that new systems and forecasting processes are important, but that talent management will improve their firm’s forecasting ability and will lead to better financial results for stakeholders. Heizer and Render (2014) discuss human resources in chapter ten of Operations Management, and supply chain in chapter 11, although they do not integrate the two topics such as this article does.

References

Chase, Charles W. (2015-2016) The Importance of Product Segmentation. The Journal of  

Business Forecasting, 34, 4, 36-40. Retrieved from https://search.proquest.com

Chase, Charles W. (2016) Forecast Accuracy Has No Impact on Inventory! Really? The Journal of

Business Forecasting, 35, 1, 26-29. Retrieved from https://search.proquest.com

Heizer, J., Render, B. (2014). Operations Management: Sustainability and Supply Chain

      Management. Upper Saddle River, NJ: Pearson Education, Inc.

Homareau, Jack. (2015) Forecasting Sales Volumes with Economic Indicators.  The Journal of    

       Business Forecasting, 34, 3 32-34. Retrieved from https://search.proquest.com

Jain, Chaman. (Winter 2006/2007) Benchmarking Forecasting Models.  The Journal

      of Business Forecasting, 25, 4, 14-17. Retrieved from https://search.proquest.com

Jain, Chaman. (Winter 2017/2018) Letter from the Editor, The Journal of Business Forecasting,

36, 4, 3. Retrieved from https://search.proquest.com

Wagner, Rich. (2016) Why External Data Are Vital to Demand Forecasts. The Journal of Business  

Forecasting, 35, 1, 30-34. Retrieved from https://search.proquest.com

Wilson, Eric and Breault, Jason. (2016) Improving Forecast Accuracy through Talent

Management.  The Journal of Business Forecasting, 35, 1, 4-9. Retrieved from

https://search.proquest.com