Demand forecasting software can usually factor in climate and seasonality, like more ice cream being sold during summer months or in warmer climates, but weather data adds a more granular time-sensitive signal.
Weather can drive demand: like heat and beach goods, rain and umbrellas, or snow and shovels. But forecasting demand for weather-dependent goods can be challenging because it can be more difficult to predict. As meteorologists say, “Climate is what you expect and weather is what you get.” There are several issues to consider with weather-influenced demand:
- When you view historical sales, you can plot the demand peaks and valleys relative to the weather signal over any period of time. But when you build a model with historical data and test it with forward-looking forecast data, you often can’t correctly predict an accurate peak in demand because beyond the general climate profile, you often can’t predict the weather forecast far enough and accurately enough into the future.
- It’s easier to forecast aggregated weather-influenced demand across a wider area rather than in a smaller, localized zone. But that’s the granular level you need to forecast for supply chain planning.
- You need to represent demand lags. Weather changes often do not immediately translate into demand. There may be a lag in sell-out—selling out of the store to the consumer. Ice cream sales may be impacted immediately, but air conditioner sales may not materialize for several days. In addition there may be a lag in sell-in, the store buying items through its supply chain.
The upshot is that improving forecast demand based on weather can be very complex, depending upon areas, products, and different expected lags in demand. It is very difficult to model statistically, because there are so many variables.
But it is amenable to machine learning, which can crunch enormous amounts of data. And with machine learning, you don’t have to make a lot of statistical assumptions on lags for the different types of demand. You input the data, and let the machine “understand” the demand latency depending on product and area.
For instance, in a recent application that involved temperature and ice cream, a consumer goods company deployed machine learning to develop a model for five European countries: Germany, Poland, Sweden, the Netherlands, and France. In the above graph, the difference between the weather-enhanced forecast and the original forecast is shaded in red where weather data added forecast value (FVA), and in blue in those few places where the original forecast was closer to the mark. As you might expect, the weather data added value during periods of increasing temperatures and sales (weeks in spring and summer) but not during the colder low season weeks. Overall, the demand forecast improvement was significant— averaging 3.8% across all five nations and as high as 5.8% in one country where local weather was the most predictable and had the biggest impact.
Machine learning can also be employed to improve more complex climate-driven demand forecasts. For instance, Lennox Residential, a leading manufacturer of HVAC systems, uses machine learning and cluster analysis to automatically identify and track seasonality patterns and trends. The software recognizes more than 200 “micro-climates” within the United States and their seasonal timing variations, and factors in weather patterns to “right-size” demand forecasts for more than 450,000 SKU-locations.
Lennox reported at a recent CSCMP global conference that with so many seasonal variations and data points it “would be virtually impossible to analyze with traditional tools and methods. However, by employing machine learning, our solution automatically sifts through massive amounts of ERP data and identifies ‘seasonality clusters’ that exhibit similar seasonal demand patterns. Refining demand trends around these clusters yields more accurate results – and, with machine learning, forecast accuracy automatically improves over time because the clusters are continually adjusted based on the latest information.”
Good weather data is becoming increasingly available. IBM recently acquired The Weather Company, partly so Big Blue customers can marry weather data with connected device data streams in the Watson cognitive cloud.
So, where do you start?
“You begin by understanding what you want to do at a business level,” says Pietro Peterlongo from ToolsGroup’s advanced analytics team. “Then you model the data, and let machine learning process it iteratively and ‘self-learn,’ while you evaluate the results. In this way, you can use weather forecasting in the same way that you evaluate causal factors like pricing and traffic—to get the best picture of demand for a particular product during a specific time series.”
Click below for a short whitepaper that describes many ways to leverage data to improve demand forecasting.
Gartner comments that interest in predictive analytics has increased recently for a myriad of factors. Supply chain organizations strive to become more predictive in managing their supply chains to take advantage of potential upsides and proactively mitigate potential disruptions. The availability of supply chain data — such as Internet of Things data, dynamic sales data and weather patterns — provides the ability to extrapolate the current environment to better understand future scenarios. They add that interest and adoption of predictive analytics has enjoyed a significant increase thanks to corresponding interest in machine learning and deep neural networks that are capable of generating more accurate predictions with little human intervention.