Pandemics. Wars. Energy crises. Unprecedented circumstances can make it hard to build accurate forecasts, especially when forecasts are based on historical data alone, explains Josh Ackerman, a Data Scientist Manager at DOW.
“It’s hard to make strategic decisions when you can’t predict the future from the past,” says Ackerman.
Leaders at DOW challenged Ackerman’s team to build a forecasting tool that incorporated more than historical data, including more data inputs and more statistical models. To get it right, Ackerman met with teams across the company to understand decision-makers' needs and expectations for a forecasting tool.
“Historical attempts at forecasting were based on statistics and correlation analysis,” he says. “The forecasts were pretty good, but if business users didn’t understand them or vet the indicators in the forecast, they didn’t use them.”
The old forecasts used black-box models with ensemble algorithms and internal indicators. Ackermans says that because decision-makers couldn’t explore the data's inputs, they didn’t fully trust the results.
Ackerman shared DOW's forecasting journey at SAS Innovate in Dallas during a manufacturing breakout session.
Building forecasts that decision makers will use
The DOW data science team built a new forecasting system that incorporated industry-specific market intelligence for customers and products, along with high-level macroeconomic indicators such as crude oil prices, unemployment rates, long-term interest rates, and the consumer price index.
“A lot of the data collection had to be intentional,” says Acerman. “We engaged with the business users, vetted the models and helped build ownership with the business users.”
The new system, built with SAS, is a white-box system that tells users which indicators are used for each forecast and shows how the inputs relate to demand patterns. Decision makers can ask questions about the data as they explore the forecast.
“When meeting with users, we ask, ‘Are we using the right indicators? Do they have the right relationships with demand that make sense to users? What does accuracy look like over the next few months?’" Then the team compares the forecast with univariate results to double-check accuracy.
Ackerman calls the ARIMAX model a SAS forecasting workhorse and says it’s used for trend analysis, cycle-seasonality analysis, cross-series analysis and time-series analysis.
The previous system had 100s of forecasts at the region and industry level, and today’s system produces 1000s of forecasts monthly at the region, industry and product family level. It provides long-term market forecasts for 4-14 months, horizon forecasts for aggregate planning and near-term forecasts to monitor and focus on execution of demand plans.
Forecasting provides millions of dollars in value
Results for the business include fewer stockout issues, reduced discount sales and more products reaching the right places at the right time.
When asked about ROI, Ackerman says he estimates the system has already provided ten times the value for the cost of implementing it.
“Reliable statistical demand forecasting offers big benefits,” says Ackerman. He estimates that improvements in product availability, reductions in inventory and shorter time-to-build for monthly forecasts will generate millions of dollars in value each year.
