More and more CFOs — and companies — are applying predictive analytics to boost planning and forecasting accuracy and solve an ever-increasing range of business problems.
One reason: The barriers to using predictive analytics tools, which employ statistical models to make forecasts and projections and uncover key business drivers, are being lowered. Falling costs and improved technology mean the tools are gaining wider acceptance as more companies experience success with them.
Many predictive analytics methods have been available for years, but only for companies with the resources—like a staff of data scientists, statisticians, and programmers. The time to commit to a considerable project was another prerequisite, says George Mathew, president and chief operating officer of Alteryx, a self-service data analytics firm. But software developers are now selling predictive analytics tools that are accessible to finance executives, as well as other company managers and staff who aren’t programmers or data scientists but understand the problems that analytics can solve.
“Many of the users understand the quantitative analytics surrounding this work, but they’re not programmers, so they’re not going to necessarily program in Python or whatever it might be to create an accurate predictive model,” Mathew says.
More Robust Models
Finance executives are using predictive analytics for planning and forecasting, risk management, and compliance and controls, like detecting anomalies such as possible money laundering by supply chain partners, Mathew says.
With the increasing availability of predictive analytics techniques, such as machine learning, CFOs can also build more robust models to forecast revenues or P&L, says Anshuman Mishra, data scientist in financial services at Alpine Data. The models can inject more dependencies into the analysis, factoring in changes in financial market indices, cost of capital, or inventory pricing. The models can also extract data from pubic company 10-Ks and bankruptcy filings.
“You get access to much more nuanced data, nuanced analysis, when you apply machine learning,” Mishra says.
Better technology is making the barrier to entry for running machine learning models on very large volumes of data quite easy, says Josh Lewis, Alpine Data’s vice president of products. Today’s data storage platforms, such as massively parallel processing databases, allow much more sophisticated computing than earlier databases that didn’t allow such manipulation.
“Storing large amounts of data has always been pretty feasible, and now we’re getting really good as an industry at sophisticated computing on top of large amounts of data,” Lewis says. “Companies like ours are creating tools to distance the end user from the complexities of those computations and make it something that’s really visual and enables citizen data scientists.”
While costs currently restrict full machine learning initiatives to big projects with high-value cases, there are opportunities as those costs come down, thanks to open-source technologies and more nimble tools, Lewis says.
“Those lower-value use cases that might be moving the needle by tens or hundreds of thousands of dollars, instead of millions of dollars, are now cost effective from an effort and return-on-investment standpoint,” he says. In the world he foresees, instead of two dozen use cases for machine learning, the landscape will be unlimited, including use cases that were never thought of before.
How are companies deploying predictive analytics? The following stories demonstrate the power of predictive analytics to boost bottom lines.
A Fortune 500 financial services company with about 1,500 directly employed agents selling annuities, life insurance, and related products used predictive analytics to determine the common traits displayed by its best customers. The company was a customer of Birst — a cloud-based business intelligence and analytics provider — which did not identify the client. The customer’s goal was to target its sales efforts toward prospective customers with the same traits, says Pedro Arellano, vice president of product strategy at Birst.
Birst helped the company launch its first predictive analytics application — a classification analysis, sifting through client data to find out the best targets to call, along with which specific products each prospect was most likely to buy and the dollar value of that opportunity.
The results: Revenue from the sales agents increased 20% in the first year and 10% to 15% per year in the following years. The success ratio for sales calls doubled, from one sale per 10 calls to two sales per 10. Client retention rates also improved.
The success with the project opened doors to other areas of the insurance company that wanted to see how they could apply predictive analytics, such as the company’s human resources division.
EviCore, a medical benefits management company with $1.5 billion in annual revenue, used machine-learning software from Alpine Data to improve the efficiency of its claims pre-approval processing. EviCore processes about 2.5 million patient transactions per month on behalf of its insurance company clients.
Before employing the predictive analytics software, eviCore would rely on its army of about 800 nurses and doctors to decide on claims preapprovals, says Josh Lewis of Alpine Data. The staff sifted through a large volume of data, ranging from patient records and details on the medical procedures ordered to insurance provider requirements and codes.
The machine learning software took the large data set of prior claims pre-approved or denied by humans, deduced general rules that would lead to those classifications, and applied those classification rules to new claims. It also incorporated the new claims outcomes into the ever-growing data set and continually updated its rules for pre-approvals, Lewis says.
When the software scored claims as extremely likely to be approved, those claims skipped the human processing step. That allowed eviCore to increase its ratio of claims processed to headcount as the business expanded, so it didn’t need to hire as many new employees to sustain the business.
A third example: The family-owned Oberweis Dairy used SAS predictive analytics software to solve problems with customer attrition and identify expansion locations.
Bruce Bedford, an Oberweis vice president who heads its analytics initiatives, says the first challenge the Illinois-based company attacked with predictive analytics was a customer attrition problem in its milk and dairy products home delivery business, which is about one third of the company’s overall sales.
Oberweis used SAS analytics software to run a “survival analysis” of transaction data from the week-to-week home delivery orders. (A survival analysis is a set of methods for analyzing data where the outcome variable is the time until the occurrence of an event.)
The analytics determined that a promotional effort for home delivery service to new customers — which included free delivery for six months — had a substantial impact on customer retention. Waiving the $2.99-per-week delivery fee for six months turned out to be a bad idea, because customers often dropped the service once the fee appeared on their bill after 27 free deliveries. Oberweis changed the promotion from six months free to a reduced fee for a longer period, and the sharp attrition rate at six months vanished. Retention rates for those new customers also improved 30%.
“That’s a counterintuitive idea, and many people were extraordinarily resistant to doing it, because they thought ‘My God, how can you charge more and get a better result?’” Bedford says.
Oberweis also applied predictive analytics to a geographic expansion issue with its home delivery service. The company has eight distribution points — three in the Chicago area and one each in Milwaukee, St. Louis, Detroit, Indianapolis, and Virginia — and plans to add two more sites in the next three to six months. Oberweis wants to expand where customers are most likely to sign up and stick with the service. The closer home delivery customers are to their distribution point, the more likely they are to remain loyal.
The company had used the analytics software to maximize the cost effectiveness of direct mail marketing to sign up new home delivery customers for its existing distribution points, drawing on demographic data for tens of millions of families on U.S Postal Service carrier routes to determine which ones should be targeted.
The direct mail model was effective enough that the company could buy a two million-piece mailing and know with a high degree of certainty what kind of return it would bring and when, Bedford says. So Oberweis took the direct mail model that was working for the markets it was already serving and applied it to potential new markets.
“The CFO cares about this decision because the minute you tell him you need to go to a new market, he’s seeing all kinds of capital expenses,” Bedford says. Those include buying or leasing property, buildings, trucks, and equipment, and hiring and training staff. “There’s a lot of expense tied to that decision, so immediately the question is: ‘If we’re going to expand, where do we put a building?’”
The analytics model is driving the ongoing expansion process, including where to put in bids for properties, where to recruit from, and where to locate trucks for delivery, Bedford says.
What’s Driving Losses?
Finally, a large insurance carrier was losing market share to local insurance companies, and its backward-looking descriptive analytics wasn’t finding the drivers behind those losses. The company hired Protiviti, a business consultancy, to help it understand the customer loyalty issues using predictive analytics modeling.
The analysis started with 5.4 million rows of data, including customer product, revenue, and distribution data, and 69 variables. Techniques for eliminating irrelevant data then cut the total to 1.2 million rows and 21 predictors, says Shaheen Dil, head of Protiviti’s advanced analytics practice. Protiviti did not reveal the name of the insurance company.
With additional regression and machine learning techniques used to test the performance of various models and identify the top five predictors, the analysis generated a management dashboard that allowed executives to make decisions in real time, Dil says. The model could predict with 81% accuracy whether customers would leave or not, identifying 14 key predictors in product, customer, and distributor categories.
Protiviti also developed recommendations for improvements, like focusing on the correct age groups; refining distribution and cross selling; and reviewing pricing.
“Every now and then we’ll find something that the client had not thought of, and they wouldn’t have known about it without looking at the big data,” she says. “The results are not always what you would expect.”
Two factors seem to be driving the recent examples of CFOs using predictive analytics, says Mathew of Alteryx. One is the technology available that allows companies to draw from multiple outside sources of data, not just internal sources. The second factor is the increasingly strategic demands on the CFO, who now is expected to understand the future as well as report on the past.
“CFOs have to bring more sophisticated, predictive, prescriptive, and descriptive analytics into the fold, because that’s what their compatriots — the CEOs and the boards of directors — are expecting,” Mathew says. “Because of that, I think there’s a much more focused CFO function that has higher-level capabilities, particularly on the analytics side, than we’ve ever seen before.”
Keith Button is a freelance writer based in Valley Cottage, New York.
The Analytics of Ice Cream
How was weather affecting Oberweis Dairy’s store sales?
Another issue Oberweis Dairy uncovered through its predictive analytics was the weather driver for ice cream sales from its convenience stores. Prior to incorporating analytics about seven years ago, the company would document the prior day’s precipitation and temperature on its daily sales reports for the dairy stores, but the correlation was unknown.
“The problem was, nobody could tell you what it meant,” says Bruce Bedford, an Oberweis vice president. “Nobody could tell you whether changing temperature by one degree had some impact or no impact on the prior day’s sales. They knew that sales were related to weather and they knew weather was somehow important in affecting sales, but they just didn’t understand what it meant.”
One of the first issues Bedford tackled with predictive analytics, with the help of Protiviti, was to create a model that allowed the company to relate historical weather patterns to historical sales data.
The analysis revealed that sales weren’t linked just to temperature and rain, but more significantly driven by dew point, which reflects the comfort level at a given temperature. Scientifically, the dew point is defined as the temperature below which water droplets begin to condense and dew can form.
The dew point connection led Bedford to do more modeling based on dew point, and now bonus-pay plans for employees are based not just on same-store sales measures, but weather-adjusted same-store sales.
“Dew point is a factor in ultimately compensating people for the performance of stores, so it really went from a very unsophisticated look to a very sophisticated look,” he says. – KB