skip to Main Content

BLOG

Weather-Influenced Demand Forecasting in 2020: 6 Key Observations
Weather Influenced Demand Forecasting Blog

It is 2020; we now have reasonably accurate short-term weather forecasts. NOAA, the National Oceanic and Atmospheric Administration in the US, reports that the five-day weather forecasts are approximately 90% accurate, 80% for the next seven-days, establishing an acceptable level of accuracy for this crucial external input for supply chain planning, especially short-term demand forecasting.

Kudos to those writing about the way weather influences consumer demand. Some articles go back a decade or more. However, things have changed. We just recently had the critical components in the game and they changed the way we think about weather-influenced demand forecasting:

 

  • better accuracy of the weather forecasts,
  • wider and cheaper accessibility of the weather data at the local and international levels,
  • a critical mass of practical AI/ML algorithms to complement traditional stochastic methods and
  • the scale of cloud computing, especially hardware-based parallel computing, to build practical models.

Based on our recent work with a Fortune 100 CPG company and a very large small-format retail chain that sells many of that company’s products, we gained an enhanced understanding of weather-based demand forecasting in the commercial world. The process was by no means easy. Key observations to share:

 

1) Vantage point matters

Both a CPG company and a Retail chain can add weather as an external factor in their forecasting workflow, but they will have to operate with different product resolutions and forecast horizons, resulting in vastly different predictive models. For example, as a Retailer, you need to know what specific products to put on the shelf now, if you know that, in a couple of days, a winter storm will hit where 10% of your stores are located. On the other hand, as a CPG company, you need to know how much ice cream and in what mix to build this month in advance of the hot summer season for the upcoming Retail orders.  Also, the Retail chain has a much richer set of data such as point-of-sale details, availability of competing products on the shelf, pricing, and promotions on substitutable products that have immediate sales effect. A CPG company can improve its predictive model results only if it has this type of additional data.

 

2) Location resolution matters

We instinctively know -and can statistically show- that you can generate more accurate weather-based demand forecasts when the forecasts are aggregated over a wider geographic area. However, for the operational planning of a supply chain, we most often need highly granular location data. That is why dimensional reduction through smart geo-clustering is important: clustering locations within which the weather is quantifiably similar. We call these the “micro-climate clusters.”

 

3) Product resolution matters

Clearly, some product demand is highly correlated to weather. The classic examples include ice cream, BBQ meat, umbrella, snow shovels, packaged water, etc. The key is to quantify this correlation for all products across different time periods and micro-climate clusters. Furthermore, products with shorter shelf-lives such as fresh BBQ meats or fresh fruit are the most sensitive to weather-based stock-outs or spoilage, which make them good candidates for quick demand-shaping actions.

 

4) Planning frequency and lead-times matter

When the time lag between shelf replenishment and sales to the consumer exceeds the crucial accuracy threshold of 5-7 days, the quality of short-term weather forecast degrades fast. Rest assured, the weather data are still useful for analyses in other time horizons: year-over-year trends, changes in month-within-year seasonality, day-within-week differences, etc.

 

5) Correlations matter

While both numeric (temperature, humidity, rain, wind speed) and categorical (cloudiness) data on weather are available, we need to account for their internal correlations, e.g., no rain on a sunny day, no snow when the temperature is 20 degrees above freezing. And then, there are correlations between demand and multiple external variables. You would agree that demand for BBQ meat in the Southeast US when college football rivals Alabama and Auburn play on a sunny October weekend is vastly different than demand for BBQ meat on any sunny October weekend.

 

6) Scalability matters

One size does not fit all when it comes to getting practical results. With our Retail client, we have had to process an average of 2,000 unique SKUs sold at approximately 9,000 locations across geography the size of California. By the time we identified weather-sensitive product clusters, weather-related seasons and the micro-climate clusters, we ended up with thousands of “weather-effect models.” There is an absolute need for the right level of automation in data processing, feature engineering, clustering, artificial intelligence and machine learning-based modeling tools, and model parameter maintenance. This work requires scale. So, if you want to step onto the field, then do not be the pray for the female lion trying to train her young on you.

In short, weather-influenced demand forecasting is not trivial. It is not a menu item you can buy in a fancy restaurant you evaluate on Yelp! It is more like 10 sessions of hard gym training with a professional trainer in Equinox. We all know that the good habits and positive effects of gym training last much longer, sometimes a lifetime.

Is this all worth the effort and investment? We will report forecast accuracy results in due time when our clients are ready to expose this work. For now, it suffices to say that we already gained significant insight from weather-based data analyses and the subsequent development of the weather-influenced demand models. Our clients continue to take tactical and operational actions based on these insights, and we are happy about that.

Love to hear your feedback and any experience you may have had with weather-influenced demand models.

If you would like to chat about weather-influenced demand forecasting or if you are curious about hearing how Solvoyo makes use of external data in planning, you can book a 15-minute discussion with me using this link.

Omer Bakkalbasi

Chief Innovation Officer at Solvoyo

Omer is a seasoned supply chain professional with more than 20 years experience in supply chain management, with an emphasis on business analytics and fact-based decision making as well as delivering high-quality software and consulting services. Omer is motivated to solve difficult business problems through innovation, information technology and automation while always focusing on the end result of delighting customers.

LinkedIn: Omer Bakkalbasi

Integrated Assortment Planning: A Vital Capability for Retailers

In today’s changing market dynamics, the retail business is getting more competitive by the day.…

Coronavirus & CPG
Coronavirus & The CPG World: What to do now and how to prepare for the future?

At Solvoyo, we are also concerned about the COVID-19 pandemic. Our thoughts are with all…

See our next-generation supply chain planning & analytics platform in action
Back To Top