• Katarina Trbara

Making Your Demand Forecasts Less Wrong

Updated: 5 days ago


Ron Burgundy, The Anchorman on forecasts for the weekend

Let’s get one thing out of the way - your demand forecast will never be 100% right, but it can be less wrong. There are numerous tips and tricks on improving your forecasts, below you will find 4 areas of focus, but keep in mind they are arranged by complexity, meaning there is no point in deploying machine learning if you still aggregate actuals manually to get a forecast.


Demand forecasting and supply chain management are truly key for success, and every incremental improvement can have a big impact, or as Dr. Chaman Jain, Editor of the Journal of Business Forecasting said in an interview for Demand-Planning.com:


“How much you can save depends on the size of a company, as well as on how developed the forecasting process is. In my study of Consumer Products Companies, I found an average company can save $3.52 mil. for every one-percent improvement in the under-forecasting error, and $1.43 mil. in improving the over-forecasting error. This saving is only from the improvement in the supply chain.”


Information silos

A group of silos with a blue sky in the background

Recently I wrote about the usage of spreadsheets in SCM and it’s safe to assume they inherently create information silos and disconnected processes. Generally, everyone involved is also aware of the disconnect.


The other side of the coin is a common trap companies fall into in this era of digital transformation gold rush. They implement the latest technology or solution for each department, but they fail to connect all of these (expensive) data sources. We’ve seen this many times, a company might have an enterprise suite CRM, IoT warehouse management system, AI production planning solution, and no converging point for all the data. What good vendors should do is (a) tell you data sources need to be connected for the system to function properly and (b) integrate their solution with the existing systems.


From the perspective of demand forecasting, both cases produce a similar result - multiple data sources with no real-time updates on actuals and sales and marketing strategies.


In a recent study by Accenture 64% of companies “aren't seeing digital investments boosting their revenue growth at all.” There are various reasons for this, but the companies that did get a boost in their revenue growth from digital investments (a 27% boost in fact) are “more likely to have their digital platforms work and communicate well together (71 percent vs. 64 percent). They have clear rules for how their information technology (which supports enterprise planning) and operating technology (which controls manufacturing and operations) should work together.”



Real-time planning and forecasting


Closely related to the first point - not only do all systems need to communicate, but they also need to communicate fast. The ideal state is the real-time, or near real-time data updates and re-forecasting based on new inputs, or as Nicholas Nassim Taleb explained it:


“If you ever do have to heed a forecast, keep in mind that its accuracy degrades rapidly as you extend it through time.”

With real-time availability of data you can forecast shorter periods with higher accuracy, and improve the resilience of your supply chain to any disruptions. The infamous Bullwhip effect can be predicted and limited with up-to-date information from all stakeholders. Reporting is no longer a tedious process that takes weeks and results in analyses of already outdated information. It helps companies shift from reactive decision-making to proactive.



Machine learning


Machine learning is an application of AI used for analyzing large volumes of data, recognizing patterns, learning from them, and continuously re-training the model. It has been applied to forecast demand since the ’80s, but the large-scale application was delayed due to scarcity of data, limited storage capacity, and low computing power.


When compared to traditional statistical methods ML differs in several ways:

  • Versatile data sources: ML can use internal (ERP, POS systems), external (weather data, macroeconomic indicators), structured and unstructured (social media, marketing campaigns, videos) data sources to generate a forecast

  • It requires a low volume of manual work and a high volume of data

  • High maintenance complexity and technology investments


In this paper, researchers analyzed the application of machine learning on Demand and sales forecasting in various industries (on-demand ride services (Uber, Lyft), clothing industry, fashion products, and even new book sales forecasting) and compared it with traditional methods. ML method outperformed traditional methods in all instances, with one exception where a combination of the two provided the best performance. However, they did point out that deploying ML may not be the best choice for every company. The process requires big investments in data storage and processing capacity, and employee qualifications. SME’s operating in relatively stable markets can delay it a bit longer and rely on traditional methods and their historical demand.


“In short, a company should favor ML rather than traditional forecasting methods for D&SF when its economic environment truly requires a digital transformation, and when the enterprise can gather the resources required to take on the challenge of Supply Chain Analytics.”


Forecast Demand Variability


In traditional or deterministic forecasting, we create a forecast which consists of single values through time series. Since this type of forecast is almost never correct, everything outside the values is considered a forecasting error which is evaluated through various KPIs.


However, there is a growing trend among industry professionals of shifting this paradigm and predicting demand variability. In a nutshell, with this method you are predicting a range of possible outcomes and a probability for each one. This type of forecasting provides you with 4 data points: best case, worst case, the range between them, and their probability.


Or as this Stefan de Kok describes it:


“Consider for a moment weather forecasts. A few decades ago the weather forecast for tomorrow would be it either rained or it did not rain. Nowadays weather forecasts provide a probability of rain. If you were planning an outside activity knowing there was a 40% chance of rain you would likely make a backup plan that was inside. That 40% provides you more information for you to make better plans that is not present in the exact "no rain" prediction.”

Although this method is more refined it is still based on and limited by historical data, and it can’t help with Black swan events like natural disasters, accidents, or political turmoils. For these situations, the best approach is practicing scenario planning and having an action plan when (if) the scenario plays out.




Farseer dramatically shortens planning cycles, improves planning accuracy, and eliminates manual work. We combine the power of AI with human knowledge, for the most accurate operations and financial plan.


If you would like to maximize revenue, optimize supply and improve your Cash-Flow, sign up for Farseer Demand Planning now.